Predictive Signals

Out of Sample Strategy Performance Update

2011-08-09T13:50:00.008-04:00

A little while ago, we published a blog post on a trading signal we've developed internally based on media analytic data. In May, we launched a live version of the components of that signal, as a feature of the Recorded Future API. Our customers can pull this data directly from the API at 3:30pm, giving them time to trade before the equity markets close at 4.

Taking the same strategy we presented earlier, and using the live data as it was available to our customers at 3:30, we have rolled our backtest forward, and looked at the performance of this strategy over the last few tumultuous months. Between May 13, and August 5, this strategy returned 10.4%, while the market lost 9.9% of its value. These returns have been fairly consistent, and turnover has been similar to what we saw in our original backtests. The results are plotted above.

Of course, this is a short time window - encompassing just 59 trading sessions, and we haven't taken into account trading costs in this analysis. Still, we find these results encouraging and will continue to look for other sources of long-term signal in our studies going forward.

If you'd like to learn more about the Recorded Future media analytics API, contact our team.

Factor Modeling Media Analytic Data

2011-03-04T16:25:00.005-05:00

At Recorded Future, we’re scouring the web for predictive signals in online content. Previously, we’ve covered our efforts at complex event modeling, and liquidity modeling using news flow information. Publicly, we’ve also touched briefly on some of our returns modeling - we’ve seen instances of particular blogs that seem to have superior predictive power in terms of their ability to write about stocks that will outperform.

Recently, we’ve expanded this approach to build a whole-market factor model that uses media analytic data to predict excess returns. Using aggregate data for the S&P 500, which is available to our API customers, we’ve built a number of factors that are derived from online sentiment and momentum of S&P 500 constituents that show statistically robust predictive signals of market-relative returns over a 1-day to 1-week investment horizon in a time-series cross-sectional modeling environment.

Factor Examination
Let’s take a look at one such factor, which is based on sentiment and momentum. If we take this factor, and break it into deciles by day and then construct portfolios for each decile, we see the following cumulative continuous returns in these portfolios. We’ve included dividend-adjusted returns to the SPDR S&P 500 ETF (SPY) as a benchmark in bright orange.

You can see quite clearly that over the last two years, our top decile (in orange) has outperformed all other deciles in a fairly consistent manner. Meanwhile, the bottom three deciles (the three darkest shades of blue) have underperformed all other deciles, as well as the market. One thing to note is that this relationship is not strictly linear. For instance, our 2nd, 3rd, and 4th place deciles actually fall near the middle of the returns distribution, which may have something to do with the construction of this particular factor.

If we compare the portfolios to the performance of the S&P 500 over this period, we find that the portfolio in the top decile has a Beta of 1.08, assuming a risk free rate of return roughly equivalent to that of T-bills over the period. It has a statistically significant annualized (continuous) Jensen’s alpha of +16% over the period. When we examine the bottom two deciles under the same assumption, we see that they are high Beta portfolios (1.37 and 1.34, respectively), but with statistically significant and negative alphas, at -42% annually, and -26%, annually. As you might imagine, constructing hedged portfolios out of the securities in these deciles provides some possibly compelling trading strategies.

If you’d like to experiment with this approach yourself. We’ve made some R code available on our Google Code site which will pull in market data, Recorded Future data, and perform this sort of decile analysis on a factor of your choosing. You’ll need a Recorded Future API token to pull that data.

Soon, we’ll discuss the inclusion of a factor like this into a portfolio built using other factors based on Recorded Future media analytic data, and find out whether a portfolio like this can stand up to trading costs, and evaluate its performance in an out-of-sample context.

Turning Online Media into Big Data for Quants

2011-03-03T20:12:00.021-05:00

We recently hosted a webcast discussing applications of the Recorded Future news analytics API for quantitative finance, and a big thanks goes out to everyone that joined us. The original presentation can be viewed here and slides from the session detailing how we turn online media into actionable data as well as several case studies are below:

Recorded Future News Analytics for Financial Services

Live Webinar: Recorded Future for Quantitative Trading

2011-02-17T15:24:00.006-05:00

When: Tuesday, March 1 at 10am EST
Where: Web conference (register here)

Join us on Tuesday, March 1 at 10am Eastern time for a webcast introducing how you can apply our news analytics API data to quantitative investment and trading strategies.

Recorded Future converts the real-time stream of news, niche, and other online channels into the only source of past, planned and speculative events on the web. These events range from corporate and government announcements to discussion of what might happen in the future — speculative events. This temporal analysis is the focus of Recorded Future, the analytical tool-kit we’ve developed, and the API that’s available.

The live session led by our Chief Analytic Officer Dr. Bill Ladd will feature modeling experiments showing how Recorded Future data can be used to formulate predictive models of liquidity and volatility as well as returns around “future” events.

We’ll also provide an in-depth introduction to Recorded Future’s temporal data including how we use computational linguistics to extract and index events, entities and related statistical measures from online media to create a robust data set ripe for generating innovative trading strategies.

Register for the March 1 event!

Detecting and Profiting from the Future

2011-01-30T18:53:00.014-05:00

Recorded Future identifies and collects discussion of events scheduled/speculated to happen in the future, and we want to find ways to incorporate this information into investment strategies. We’ve already seen some predictive power with these “future” events and are in the process of "pulling it apart." Specifically, we’ve investigated the market impact of our “future” events and observed some interesting behaviors.

When we look at market returns in the 5 days before and after a forecast event, we see a slight rise in returns before the event followed by a drop after the actual occurrence of the event.

Since these events are known ahead of time, it is initially surprising that there is any price movement at the event.

Drilling into the data a little further, we looked at just the most scheduled events: earnings calls. We see that when the future event is an earnings event, on average there is a ~25bp rise before the event followed by a ~25bp drop after the event.

This was a surprisingly large movement and may fall into our crowded investment thesis noted in an earlier post. When the earnings call is coming, there is a lot of attention on it, and then that attention dissipates. Our analysis suggest that cumulative returns on average follow the same pattern.

In contrast, for all other forecast events in our system there is essentially no movement before the event and a ~10bp drop after.

These events seem to drive no activity beforehand but are predictive of a slight drop afterwards. Now, this is averaged over 16000 trades and indicates a significant relationship. We will continue to drill down into subsets of events to find additional investment opportunities.

In this analysis, we are building on two earlier blog posts where we looked at forecast events and event studies. Our forecast events occur when an event is reported to occur after the publication date. For example, “The Verizon iPhone will be available to all on February 10th.” was published at mobilemarketingwatch.com on January 11.

We collected about 20000 of these events for S&P 500 companies over roughly the last two years and found the above patterns by looking at the average market adjusted returns for these companies in the days leading up to and after reported events. Additionally, we saw volume increases in all three analyses ranging from an increase in 3 standard deviations from normal in the earnings call events to a tenth of a standard deviation above average for the data without earnings calls.

Why is there the average drop after the event for non-earnings related events? Is negative information being withheld at announcement? Is speculative and forecast related news typically negative? Watch this space for further investigation of data from our news analytics API.

Measuring Crowded Investment Strategies with Online Media

2011-01-10T14:48:00.006-05:00

Recently at Recorded Future, we have been experimenting with applying our sentiment scoring methodology to measuring the level of other concepts communicated in web content. Some ideas we have have been playing with include deceit, fear, and uncertainty. Outside of emotive language, we have also looked at capturing the level of chatter around a particular technology or business construct.

In particular, we recently developed a score to monitor the level of chatter around the concept of “momentum investing,” an investment style that has been in and out of favor with the media and the market over the years. We then applied this scored to our content and plotted the results over time. As a comparison, we look at the performance of the Monetta Fund (MONTX) a Mutual Fund that follows a momentum investment strategy.

Our theory, before seeing these results was that we would see a positive correlation between the performance of a momentum investing strategy and discussion about it online. However, as you can see in the chart above, the two metrics are generally inversely correlated for the time period in question. For 2010, the correlation of monthly changes in the metrics was -0.56. As chatter around momentum investing declines, the $NAV/share of the fund rises, and vice-versa. This finding makes economic sense, if you look at the market from in an ecological framework. To quote David Merkel of alephblog.com, “Many strategies are competing for scarce returns. Often the best strategy is the one that has few following it, and the worst one is the crowded trade.” Is there a suitable proxy for the “crowded trade” based on online chatter? Stay tuned for more research in this space.

Seconds Away From News Analytics

2010-12-06T17:58:00.006-05:00

A little while ago we gave an example of how to get access to our News Analytic content using R. While this was pretty straightforward, we wanted to find an even easier way to use the Recorded Future API. I’ve put together an example spreadsheet that loads requested Recorded Future data live into a Google Spreadsheet and then combines that data with historical finance data from the Google Finance API. The result: a spreadsheet that will populate itself with media analytic data and stock market data straight from the web. Just enter in a list of stock tickers, a date range, and a Recorded Future API token - and within a few seconds you should have plenty of data, ripe for analysis.

What can we do with it? I’ve linked a “motion chart” to the spreadsheet. After switching around the chart type, I’ve set stock price on the y-axis, time on the x-axis, and color-coded my stock prices according to momentum. I see some interesting days of high momentum, particularly for Intel. One of these seems to be focused on August 5, 2010 - the day the FTC won an anti-trust settlement against Intel.

I’ve taken advantage of Google Apps Script to write a script that picks up spreadsheet data, runs queries against our API, Google Finance’s API, and does some processing to merge the results. Data is then put in a seperate spreadsheet “Stock Data”. The motion chart updates when the contents of that sheet change.

If you’d like to put your own token in the spreadsheet, go to File -> Make a Copy, and the sheet will be editable in your Google docs storage area. To see the code that runs when you click the “Run!” button, you can then go to Tools -> Scripts -> Script Editor.

By the way, if this kind of data is interesting to you, we have it in bulk for API customers for the S&P500 and the Russell 3000 - so you don’t have to manually enter hundreds of tickers!

Buy the rumor, really, buy the rumor

2010-12-02T14:26:00.000-05:00

We ran an event study to look at stock returns following high momentum high sentiment days for the S&P 500 over the last 21 months or so. And while we did see an increase in market adjusted returns after the event, we saw much more dramatic returns before the event.

Lets take a closer look. Specifically we defined our events to be the 5 highest momentum days with high positive sentiment for each company. Momentum and sentiment are metrics we derive from online content as we harvest it and in this case, we aggregate these metrics for individual companies. In the plot we are looking at the average cumulative returns for the roughly 2000 events starting 20 days before the event and ending 20 days after. The large jump at day 0 is the return from the close the day before the event day to the close on the event day. Not surprisingly, there is a large jump on the event day. Additionally, there is a small rise after the event day. Specifically, about 28 bps in the 5 days following the event day (p=.03). Maybe enough to cover trading costs. Maybe not.

However the bulk of positive return associated with the event occurs before the event happens. Given the large jump on event day, its clear that the information isn’t completely priced in before the event occurs. But someone was buying well before the events. Someone who had access to the rumor.

These types of event studies are pretty straightforward with the Recorded Future content. Extract the data from the database, define the event and test. The opportunities here are to define the news analytic events that allow us to get in earlier on price movements like the one above. Instead of waiting for the 30bp available after it happens.

Event Studies with Recorded Future

2010-12-02T14:08:00.004-05:00

At Recorded Future we are harvesting events from online content. Given that event studies are a staple of financial analysis, it was really just a matter of time before we investigated our events in an event study framework. We've built this In R on top of our Web services API and will eventually be placing the code in our Google Code Site. I’ve got a number of posts about specific studies but wanted to put out some technical details as a stand alone posting. This post is really designed for people who want a look under the hood of how the event study results in other posts were found.

There is one nomenclature detail to address right away. The term “Event” in event studies may not be synonymous with individual “events” we observe in news data. The “Events” that are the basis of event studies are complex events defined from one or more observed news analytic “events” retrieved from online content and associated derived metrics. Going forward, the distinction between the two uses of the word event should be clear by context.

To do an event study, we start with a selected market basket of equities and a time range. Next, we define an event criteria and search our databases for instances of that event for the selected companies. Event definition is a creative step in which any of the content we collect can be used to define a complex event: sentiment, momentum, event type, event time, event attributes, source, novelty, etc. Once we’ve identified the events, we assign a date to them, generally either an event time for one of our observed events or a publication time.

After the events and dates are defined for a set of stocks, we can look at the market basket adjusted returns for those stocks -typically close of event day to close of the following day. The market basket adjusted return is simply the difference of the daily return for a specific company vs the unweighted average return of all of the stocks in the market basket for the day. Abnormal returns are these market adjusted returns associated with events and normal returns are the rest of the population of market adjusted returns not associated with events. By definition, the total population of abnormal and normal returns will average to zero. The purpose of the event study is to see if a specific approach to defining events yields event associated returns that are statistically different from zero.

We look at a number of statistical tests to make this assessment. We generally look at a collection of one and two sample t-tests as well as non-parametric versions of these tests. The one sample t-tests of the abnormal returns give an estimate of the significance of the abnormal returns when compared to zero. We also look at the one sample t-tests of all of the returns not associated with the event to ensure that these normal returns are still not statistically different than zero. The two-sample t-test between these groups can be used if the normal returns are non-zero for some reason. We look at the non-parametric versions of these t-tests as well to ensure that findings from the t-tests are not driven by outliers.

As well as returns, we can also look at abnormal volume and volatility associated with events. Daily volume is easily obtained from typical online Finance sites like Yahoo!. We estimate volatility as the log of the ratio of the daily high over the daily low for a stock. In order to put these volume and volatility measures on a comparable scale across companies, we normalize each daily value by subtracting the mean and dividing by the standard deviation of the value for the specific company over the trailing 240 trading days. The resulting value, or Z-score, is essentially the number of standard deviations a given daily volume or volatility is above or below the average level. While the average return is guaranteed to be zero across all companies for a given day, and thus across all days, this isn’t true for volatility or volume. In some cases there are overall market trends of these values both for individual days and over time. In these cases, it is the two sample t-tests that are the basis of findings of significant changes in volume and volatility.

There are several ways to visualize results from an event study. In a daily return chart, we set the event at time zero and look at the average returns across events for the individual days around the event time. In the example above we are looking at days after the event and we could easily look at days before as well. Alternatively, we can look at cumulative returns as well where we are summing the daily returns starting from a specific starting point. In the example below, we are looking at the cumulative returns starting from the first day after the event.

The dotted lines in the two graphs are the confidence limits for the associated data. In the daily return plot, it is only the 2nd day individual return that is statistically different from zero. In the cumulative returns plot, the total return becomes statistically different than zero at day two and continues to be significantly different than zero for the next three days.

There are a number of choices to make in doing these kinds of studies and this post simply lays out the approaches we’ve taken, Feel free to contact us directly if you have more detailed questions or comments.

Clean Tech vs. Mobile Media Buzz - Introducing Recorded Future’s Custom Linguistic Scoring Feature for API Users

2010-11-29T18:45:00.011-05:00

At Recorded Future, we employ cutting-edge computational linguistic technology to organize structured and unstructured content from the web into instances of content that reference particular entities and events. Once we have this information organized across dimensions of entities, events, and several time dimensions, we calculate several metrics based on this restructured content. These metrics include measures of linguistic properties of the content itself (positive and negative sentiment) as well as measures based on a global view of the data (momentum).

Lately, we've helped Recorded Future API users to apply their own custom linguistic scoring approaches based on a collection of interesting words and phrases to our web content. In effect, this allows users to score Recorded Future's database of web content according to their own belief of what is interesting in that content.

Let's take a look at a concrete example of how this might be used in practice. Let's say I pose the following question: Does chatter online around particular technologies predict stock returns for companies focused on building these technologies? Using Recorded Future's new custom linguistic scoring feature, I'll show an approach we might use to answer this question.

Let's start with our experimental setup. For the purposes of this post, I'll look specifically at the discussion of clean technology vs. discussion of mobile technology. In order to use a custom linguistic scoring metric, I first need to determine an appropriate basket of words or phrases to score a particular text fragment for its discussion of each of these topics. I developed one basket of words for clean tech, and another for mobile.

I could have picked a few words out of thin air ("apps","CDMA","android","iphone") to decide what words I should use to describe these technologies, but in this case I took a slightly more sophisticated approach. I pulled a corpus of company descriptions from CrunchBase, where the companies have been classified by industry. I calculated frequencies of words used across descriptions of all technologies in the corpus, then calculated the frequencies of words in just the descriptions of mobile companies. I then rank words by difference in frequencies between the "mobile only" corpus and the general corpus and select the 100 highest ranked words. I follow a similar approach with the clean tech corpus.

The idea here is that words that appear with abnormal frequency in the descriptions of companies in specific industries are going to be representative words unique to descriptions of companies in that industry. To give you an idea of what this exercise yielded, let's take a look at the highest scoring words in the two industries.

Clean Tech	Mobile
renewable	application
electric	text
gas	location
carbon	sms
green	devices
clean	applications
water	service
systems	phones
technology	iphone
solar	wireless
power	phone
energy	mobile

Now, we have everything we need to score some text for its level of reference to words related to these industries. So, now what data do we want to score? And how can we take this scored content and turn it into something meaningful?

Since we're exploring a link between this language and stock prices, I took all references to S&P 500 companies (entity occurrences) in our database over the period January 1, 2009 to October 17, 2010 and applied the scoring metrics to this text. I then took the average score assigned to every text fragment for each day in this date range to come up with an average score per day for each metric. The result is an aggregate level of chatter around each topic in text that references S&P500 companies over the time period. Below, you can see a chart comparing these two metrics:

As you can see, the level of chatter around clean tech has remained flat over the period, while mobile is on the rise. Now, how can we test whether there's a relationship between these metrics and market activity?

One way to do this might be to look at venture capital deal flows by industry. Another way might be to go to the public markets and take a look at the comparison between these two metrics and a couple of industry-specific ETFs. In this case, we look at the Claymore/Mac Global Solar Index ETF (TAN) for comparison with our clean tech metric, and the PowerShares Dynamic Telecommunications & Wireless Portfolio ETF (PTE) for comparison with our mobile metric:

We see a positive correspondence between the returns of these two ETFs and the changes in their respective metrics over the entire period. We also see certain sub-periods where the returns of the ETFs are strongly negatively correlated with their metrics.

Even if these relationships are not strong, they suggest a few next steps in exploratory analysis of this data - perhaps we look at relationships between days of unusual height in our mobile metric and volatility in mobile stocks? Additionally, it looks like the level of discussion around clean tech peaked in late 2009/early 2010, just as the Solar ETF was about to take a dive. Is this more than coincidence? We'd have to dig into the data to find out.

Of course, I didn't have to pick clean tech/mobile as metrics here - I didn't have to even pick an industry. Just a basket of words used for classifying language. Could be "tightening" talk by central bankers, or deceitful language used by CEOs. I also didn't have to pick mentions of S&P 500 companies as my source text to apply scoring to. I could have chosen any entity occurrences in blogs written by my favorite economic bloggers, or around takeover rumors in the Healthcare sector.

The possibilities are endless. If you're interested in this idea, please contact us to obtain an API token.

Drawing Sparklines with Recorded Future data and Google Chart API

2010-11-23T09:57:00.007-05:00

Over at our Google Code site, we have been busy adding content and new examples of how to use the Recorded Future API in custom applications and trading environments.

We often get asked about embedding graphical news analytics data into web applications, so one recent code example is an effort to show how one might do that.

Of course, we already offer the "Embed" feature along with any of our visualizations in the Web User Interface, which allows for easy copy-and-paste of HTML code to include particular timeline or network views in any website. However, sometimes our customers want more control over which data is displayed, how big the embedded object is, etc. For these types of tasks, we turn to our Web Services API.

The code provides a function (generate_sentiment_sparkline) that takes in a ticker, date range, and API token, and returns a PNG image that offers an overview of average sentiment surrounding the ticker in question over the specified time period. At a high level, it works as follows:

Perform a lookup to find the Recorded Future Entity ID for the input ticker.
Perform a query to get two daily sentiment series for the ticker over the specified date range.
Perform some smoothing on that raw data.
Call out to the Google Chart API to generate a chart image (in PNG) for later display.

Here's an example of what the output looks like for Apple over the last 6 months. It is important to note that the visual properties of this chart are completely customizable, and this example is meant to show how to draw two barebones sparklines of media sentiment data:

We won't dig too much into the content of the chart here, but you can see very quickly some interesting visual trends in use of emotive language around Apple over the last few weeks. Of course, we can dig into the data more with our favorite analytic tools, but the ability to quickly generate images like this is potentially very powerful to our users who are interested in monitoring changes in media sentiment around a set of companies.

Live Webinar: Recorded Future News Analytics API for Quantitative Trading

2010-10-25T10:32:00.017-04:00

When: Monday, November 1, 2010 at 11:00AM EST
Where: Web conference (register here)

On Monday, November 1, we're hosting a webcast to formally announce the launch of our Recorded Future news analytics API and to demonstrate a simple modeling approach that can be set up in set up in less than five minutes! Join us to see how our API short web seminar to see how our API data is used to support quantitative investment and trading strategies.

We’ll open by detailing the pricing and plans for our API product and then provide an in-depth presentation of Recorded Future’s temporal data and web service capabilities led by our Chief Analytic Officer, Dr. Bill Ladd.

In a demonstration of the API, we’ll show the retrieval of Recorded Future data live in an R analytic environment, how to integrate with existing financial data, and an initial modeling exercise. Additionally, we’ll discuss how the API can be used for activities ranging from construction of alpha-generating signals to regime change detectors.

We will cover how, using computational linguistics, we extract and temporally index events, entities, and related measures from a wide variety of online media. This structured data identifies historical, current and expected future events as well as associated statistical measures such as momentum and sentiment.

Register for the November 1 event!

Minutes Away from Analyzing News Analytic Data

2010-10-15T15:23:00.003-04:00

The Web Services API for Recorded Future is designed to answer some fairly sophisticated questions. Some people are writing applications using our API to integrate Recorded Future content into their internal proprietary applications. This blog has posted the results of several complex analyses and a variety of code examples using both R and Python are available on our Google Code site

But even if Web Services, JSON and Python are not familiar words to you, its pretty straightforward to get data and start modeling. I have an example R program that you can get up and running in a few minutes that will retrieve average sentiment and momentum for a company as well as its market performance since January of 2009. With this data its very straightforward to start looking for relationships in the data.

It will of course be easiest to do this if you are familiar with R, but you simply need to install R and download three packages, fImport, rjson, and RCurl. Once you have an R session running,
you can enter these lines of code to download and install the required packages. (You only need to install once; in subsequent sessions, just use the library commands.)

> install.package("rjson")
> install.packages("RCurl")
> install.packages("fImport")

> library(RCurl)
> library(rjson)
> library(fImport)

The rest of the code in the file sets up an R function createDataSet (after you supply an access token for the Recorded Future API). This function accepts a stock ticker as input and retrieves Recorded Future data for the average sentiment and momentum for each day over an 18 month period. It also retrieves stock prices and trading volume for the ticker and for an ETF that tracks the S&P 500. From the retrieved data, it calculates a number of derived values including:

The difference between positive and negative sentiment
The positive and negative sentiments multiplied by momentum
The returns for the selected ticker and for the market index for the day after the news analytic metrics are calculated
The market adjusted returns of the ticker

Subscribers to the Recorded Future API can now easily create a simple news analytic data set for an individual ticker, say the ticker for Amazon.com, as follows:

> out<-createDataSet("AMZN")

yielding a data frame structured like this:

> tail(out)
Day Entity Count Momentum Positive Negative Ticker
265 2009-09-22 33328212 64 0.0803539 0.0896110 0.01840840 AMZN
266 2009-09-23 33328212 157 0.1367590 0.0645906 0.01073010 AMZN
267 2009-09-24 33328212 34 0.0782548 0.0388494 0.00448654 AMZN
268 2009-09-25 33328212 76 0.1315870 0.0913595 0.01383090 AMZN
272 2009-09-29 33328212 52 0.0822941 0.1491080 0.00000000 AMZN
273 2009-09-30 33328212 64 0.0961285 0.0641504 0.04386140 AMZN

sentiment.difference weighted.pos weighted.neg Open High Low Close
265 -0.01840840 0.007200593 0.0014791867 91.46 94.19 91.10 93.75
266 -0.01073010 0.008833346 0.0014674377 92.82 94.50 92.22 92.38
267 -0.00448654 0.003040152 0.0003510933 92.00 92.71 90.77 92.11
268 -0.01383090 0.012021723 0.0018199666 91.44 92.25 89.75 90.52
272 0.00000000 0.012270709 0.0000000000 91.96 92.33 90.10 91.72
273 -0.04386140 0.006166682 0.0042163306 92.26 94.17 91.43 93.36

Volume Adj.Close returns SPY.Open SPY.High SPY.Low SPY.Close
265 8264900 93.75 -0.034619050 107.08 107.37 106.60 107.07
266 5685300 92.38 0.014721160 107.32 108.03 105.99 106.18
267 5075100 92.11 0.002926990 106.41 106.64 104.55 105.01
268 4256800 90.52 0.017412694 104.78 105.36 104.09 104.45
272 4393900 91.72 0.005328127 106.51 107.02 105.78 106.00
273 8539200 93.36 -0.017722530 106.36 106.46 104.62 105.59

SPY.Volume SPY.Adj.Close spy.returns marketAdjustedReturns weekday
265 143126700 105.43 -0.005802632 -0.028816418 Tue
266 225947400 104.55 0.008381800 0.006339360 Wed
267 228636800 103.40 0.011060464 -0.008133474 Thu
268 204059000 102.85 0.005333346 0.012079348 Fri
272 133733900 104.37 0.003061324 0.002266802 Tue
273 254383000 103.97 0.003839882 -0.021562412 Wed

Note that sentiment is divided into individual measures for positive and negative sentiment. We find these independently interesting values and model them independently.

With this data frame created, there are many exploratory possibilities. For example, we can create a linear model of relationships between our news analytic metrics and market adjusted returns

> model.fit<-lm(marketAdjustedReturns~Momentum+Positive+Negative+
weighted.pos+weighted.neg,data=out)
> summary(model.fit)
Call:
lm(formula = marketAdjustedReturns ~ Momentum + Positive + Negative +
weighted.pos + weighted.neg, data = out)

Residuals:
Min 1Q Median 3Q Max
-0.127379 -0.012972 0.002639 0.012891 0.057185

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.019392 0.009178 -2.113 0.03631 *
Momentum 0.171907 0.081764 2.102 0.03723 *
Positive 0.319033 0.101889 3.131 0.00210 **
Negative -0.313860 0.228773 -1.372 0.17219
weighted.pos -3.435493 0.833410 -4.122 6.27e-05 ***
weighted.neg 4.197043 2.132710 1.968 0.05097 .
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.02424 on 146 degrees of freedom
(4 observations deleted due to missingness)
Multiple R-squared: 0.1495, Adjusted R-squared: 0.1204
F-statistic: 5.134 on 5 and 146 DF, p-value: 0.0002285

This model suggests relationships between the market adjusted returns and momentum and positive sentiment and perhaps a marginal relationship with negative momentum. At this point, we are just scratching the surface of the models we could build and interpret. Additionally, we could easily loop through a collection of tickers to generate a larger set of data. The point is that once you have an access token you are literally just minutes away from analyzing news analytics data in R.

Recorded Future Day Trading

2010-10-08T17:11:00.003-04:00

At Recorded Future, we often get asked how our news analytic technology can fit into an automated trading strategy. The answer: We offer a rich web service API that provides near real-time access to content as it is processed by our system. We've just posted some example R code to our Google Code site which illustrates how one might incorporate our API into this kind of strategy. At a high level, the code works as follows:

1) Load up a query which monitors for new content related to S&P 500 companies.
2) Every five minutes, poll the Recorded Future API for new content related to these companies (on the basis of the time the document was analyzed by our system).
3) If we see a new occurrence of one of these companies in source content, check to see whether that occurrence matches the following criteria:

Is this content truly relevant to the company at hand (using our new "relevance" score)
Does the content have sufficient positive (and insufficient negative) sentiment associated with it?
Does the company in question have sufficiently high momentum?
Is the company NOT already in our portfolio?

4) If the occurrence matches those criteria, we execute a paper "buy" order on the basis of current stock price. (Using near real-time quotes from Google Finance)
5) At the end of the day, get closing prices for every stock in our paper portfolio, and execute a paper sell at this price.
6) Calculate profits and losses on the basis of the trades made during the day.

Let's have a look at the results of this strategy, which was run on live data from Friday, October 1, 2010:

	ticker	trade_time	price	close	returns(%)
1	GE	09-55-00	16.48	16.36	-0.728155340
2	ORCL	09-55-00	27.37	27.24	-0.474972598
3	GOOG	09-55-00	528.17	525.62	-0.482799099
4	WFC	10-00-00	25.42	25.56	0.550747443
5	BAC	10-15-00	13.15	13.30	1.140684411
6	T	10-20-00	28.82	28.81	-0.034698126
7	MSFT	10-25-00	24.50	24.38	-0.489795918
8	AMZN	10-30-00	152.99	153.71	0.470618995
9	KFT	10-40-00	30.92	31.21	0.937904269
10	AAPL	10-45-00	283.13	282.52	-0.215448734
11	WMT	10-50-00	53.29	53.36	0.131356727
12	HPQ	11-00-00	40.86	40.77	-0.220264317
14	MOT	11-20-00	8.55	8.56	0.116959064
15	VZ	11-30-00	32.87	32.89	0.060845756
16	S	11-35-00	4.65	4.72	1.505376344
17	F	12-10-00	12.38	12.26	-0.969305331
18	YHOO	12-15-00	14.18	14.27	0.634696756
19	GS	13-05-40	147.44	147.70	0.176342919
20	INTC	13-25-49	19.29	19.32	0.155520995
21	C	15-01-14	4.09	4.09	0.000000000
22	IBM	15-51-26	135.65	135.64	-0.007371913
23	NYT	15-51-26	7.83	7.85	0.255427842

You can see that we executed 23 "buys" at various times throughout the day. Our average profits were +11bp/trade, with our best being +115bp, worst being -96bp.

This is obviously a naive trading strategy, and customers are using much more sophisticated approaches. This approach does not include trading costs, carries with it an extremely small sample size, and has no risk control parameters. The purpose of this example is to show how one could include the Recorded Future API into a live trading strategy and to make available sample code for performing these operations. Via the API, we also offer this data historically for the purposes of modeling and strategy building. Take a look at our Google Code Site for more information about our API and contact sales@recordedfuture.com for more information about getting access.

Live Webinar: Recorded Future for Discretionary Investment Research

2010-09-27T10:06:00.006-04:00

When: Tuesday, September 28, 2010 at 11:00AM EST
Where: Web Conference (register here)

Join us on Tuesday, September 28 to learn about how Recorded Future’s temporal news analytics content supports fundamental research for equity and fixed income analysis.

Our CEO, Dr. Christopher Ahlberg, will introduce Recorded Future and its temporal analytics engine followed by an in-depth demonstration of capabilities and use cases supporting discretionary investment research strategies.

We’ll demonstrate how, using computational linguistics, we extract events and entities from a wide variety of online media and organize the data according to temporal indicators.

Also, we’ll show how our user interface provides efficient navigation of historical, current and expected future events supplemented with statistical metrics such as momentum and sentiment. Then we'll discuss how Recorded Future can be used within your existing work flow to research potential investment opportunities and evaluate existing positions.

Included in the demonstration will be how to:

* Recognize investment opportunities through analysis of planned or likely future events
* Organize media data from disparate sources through a consistent and unified interface
* Visualize online momentum, corporate trends and the emergence of new technologies
* Identify relationships between firms, products, influential analysts and stakeholders
* Set alerts to monitor for corporate or industry events during targeted periods of time

Register Here

We look forward to you joining us!

White Paper - News Analytics for Quantitative Trading Strategies

2010-09-20T07:54:00.043-04:00

Recorded Future is building news analytics for large scale analysis of online media flow that spans blogs and Twitter to mainstream news to government filings. The white paper discussing our temporal analytic approach can be found here.

Although our content has broad application across many domains, we have had significant initial interest from the area of algorithmic trading across asset classes. This document will focus on some news analytic approaches relevant to this area.

News Analytics

In order to define investment strategies, quantitative investors take a variety of data streams, build models based on principles such as pair trading, mean reversion, etc. They assess these models with back testing and other historical simulation methods, and implement them in trading strategies.

Recorded Future news analytic content fits directly into that approach as an additional set of news analytic data streams that may either be modeled on their own or in conjunction with other data streams.

In some cases, the Recorded Future data streams may be explored for statistically significant relationships with market outcomes of interest and when these are found, optimized and included in trading strategies. Other approaches may simply evaluate a variety of trading strategies based on the Recorded Future data.

The point of any analysis in support of investment is to motivate a change in positions. In the end, any signals of interest, continuous, discrete, or composite will be applied in a trading strategy.

Before diving too deeply into modeling issues, it is important to consider the classes of signals available in Recorded Future content.

These can be broken into discrete and continuous data types:

Measures and metrics: Continuous Data Types

Continuous streams, i.e. momentum, sentiment, hedging, entity volume, document volume, are measured or calculated quantities that vary over time for specific events and entities.

Momentum is a measure of the “buzz” around a specific entity (person, company, place) or event type ("merger," "person travel," etc). It is based on short, intermediate, and long term levels and change in content, as well as source credibility and a number of other factors. Think of it as a “Google Page Rank” for media flow content.

Sentiment measures include metrics of the positivity and negativity of the language used in the context of entity or event while hedging is a measure of the certainty in the language describing an entitity/event. On the other hand, the simplest measurements are to simply count the number of entity instances or event instances specific to a company of interest.

These factors are essentially time series of specific metrics over time. These continuous measures can be refined (subsets) and aggregated (averages on supersets) to specific groupings of interest as desired.

For example, company record volume, sentiment and momentum can be grouped by industry, market cap, etc. These measures can also be broken down further; one could examine sentiment or company record volume from specific media source types, media topic, geography, etc.

Additionally these time series can be evaluated in different frameworks. Content can be interpreted according to the time it is published or according to the time it is made available in our system. Typically this difference is small though it can occasionally be large, for example when adding a new historical source. This choice might depend on what type of backtesting one is interested in performing.

One might also want to focus on event time. As new events are added into our system, we determine when these events are stated to occur, whether it's in the past, present or future. These event times are particularly useful in finding and analyzing predicted future events.

Event and Temporal Data: Discrete Data Types

The core records in the Recorded Future database are event and entity instances. Entities are typically companies, people or geographic locations while there are currently ~150 event types including "Quotation," "Acquisition," "Earnings Call," to name a few.

Consider an event such as a quotation from Ben Bernanke about the federal funds rate. The Recorded Future database will contain a record of specific event instances for this over time. Each of these instances is an atomic event, derived from a single observed event and can be used in further modeling. It is also possible to generate discrete events from continuous measures, for example a specific company having a momentum change of X over the course of a week.

Atomic events can be grouped together to form composite events. For example, three or more press releases and two or more insider trading events happening in the same week for a given company is a composite event. We can create a single event from a set of rules applied to atomic events. The rules for defining a composite event may be arbitrarily complex and may include partial time ordering as well as the occurrence of specific intra-relationships between atomic events (i.e. the press release and the insider trading events all correspond to the same company)

These composite events are closely related to complex events and their detection and analysis is related to complex event processing. As defined here, the composite event is the collection of aggregated atomic events and the complex event is a higher level event inferred from the existence of the composite event, perhaps significant changes occurring at a company that meets these criteria.

Signal Analysis

Modeling Market Metrics with Continuous Recorded Future Variables

Analysis of continuous data may be performed using a variety of regression approaches examining the explanatory power of the continuous data against outcomes of interest such as returns, trading volume, or volatility. Other predictors may be added to see if the Recorded Future continuous data provides explanatory power after compensating for other variables such as S&P performance (other common predictors include...).

In one such analysis posted on our blog, we looked at whether or not differences in momentum for a company were predictive of changes in market volume following the momentum change. In a regression controlling for both the previous days volume and the average volume over the last 20 days, we found a statistically significant relationship between the previous days momentum (weighted by the trailing average volume). The specific model fit was:

DVt = a*DV(t-1) + b*SMA(DV, t-1, t-20) + c*(MOt-1*SMA(DV, t-1, t-20)) + et

Where DVx is Dollar Volume at time x, SMA provides a simple moving average function on a range of time periods, MO is the Recorded Future momentum measure, and et is the error term at time t. We performed the analysis in the statistical computing environment R and the fitted model was:

Call:

lm(formula = Dollarvol.1 ~ 0 + lDollarvol.1 + smaDvol.Dollarvol.1 + smaxlMo, data = seriesdf)

Residuals:
Min 1Q Median 3Q Max
-5.039e+09 -2.215e+07 -2.284e+06 1.813e+07 1.597e+10

Coefficients
Estimate Std. Error t value Pr(>t)
lDollarvol.1 0.513193 0.003237 158.54 < 2e-16 ***
smaDvol.Dollarvol.1 0.471645 0.003817 123.56 < 2e-16 ***
smaxlMo 0.077162 0.015683 4.92 8.67e-07 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 170900000 on 72109 degrees of freedom
Multiple R-squared: 0.8539, Adjusted R-squared: 0.8539
F-statistic: 1.405e+05 on 3 and 72109 DF, p-value: < 2.2e-16

The positive coefficient for the SmaxlMo term implies increasing trading volume with increasing momentum. More details on this momentum analysis can be seen in our blog.

This example is just one possibility of how to use Recorded Futures continuous metrics as a predictor for market data. Many other news based approaches are possible.

Perhaps more significantly, these metrics can be incorporated into existing models to add additional explanatory power. If “news” is contributing noise to an existing model, incorporation of news analytic data may improve the model performance. Quantitative investors might consider the strategies that they are using today and assess the potential utility of adding news analytic metrics to existing models.

Modeling Market Metrics with Discrete Recorded Future Variables

The evaluation of a discrete signal for trading may involve deriving a set of potential trades (or non-trades if evaluating trading hiatus strategies) from the signal and evaluating the returns obtained by making those trades. Did the direction of the trade and rise/fall of the asset price agree more often than expected? What is the average return on trades made using the signal? What are the Sharpe/Sortino ratios of trades based on the trading signal? How do the returns to the trading signal do vs. the market? These approaches are appropriate for both atomic and composite discrete events.

In another example from our blog, we looked at “future” events for S&P500 companies where future events are limited to events “occurring” after publication, lasting one day or less and occurring on a trading day. We then looked to see if market volume on these days for these companies was higher than average.

We found a statistically significant relationship where volume on these “future” days were on average higher than on other days. We also looked at this for individual companies using a wilcoxon test and observed that for an unexpectedly large number of companies, future days had increased volume.

Histogram of P-Values for relationship between Future Events and Trading Volume. A disproportionate number of the relationships show statistical significance.

If there were no relationship between future events and volume, we’d expect this histogram to be relatively flat, with roughly 5% of the t-tests having a p-value less than 5%. In contrast, we see about 35% of our companies having significant differences between predicted event volume and non-predicted volume. This type of prediction of volumes may be useful if an investor is interested in the change in liquidity of a given stock over time.

In the last example, we looked at market metrics on days associated discrete events from the Recorded Future database and compared to market metrics on other days to see if the discrete events are associated with differences in these events. In the example from the previous section, we looked at whether there was a relationship between the continuous momentum metric and trading volume. It is also possible to combine these discrete and continuous variables into arbitrarily complex metrics as well.

For example, in a third blog post we looked at whether there was a relationship between company mentions in a specific financial news blog (FT Alphaville) and future market returns. Specifically for discrete days where a company was mentioned in that blog, we calculated a metric based on sentiment and momentum for that company and looked for a relationship between that metric and returns. We found a statistically significant relationship and interestingly enough, did not find a similar relationship across media mentions as a whole.

Thus far we have considered atomic discrete events. Composite events may be arbitrarily complex and it may also be useful to think of scoring them for potential relevance. Consider the world of mergers and acquisitions where there we might want to monitor 15-20 different classes of atomic events and trigger the composite event when a “critical mass” of the various events has occurred. “critical mass” could be a score which is built by applying scoring criteria to the underlying events. Perhaps the more sources report on a potential merger, the higher the score of the “merger” event is etc.

Signals like this may be assessed by a human for potential relevance rather than automatically triggering a trade. A composite event detection paradigm can provide value by tracking numerous lower level events that in themselves might not be informative, but when combined with other similar event streams might lead to a coherent signal.

Trading Strategies

Statistically significant relationships are important, but in order to actually generate profits from a specific signal, an explicit trading strategy must be specified. Based on a given signal, there will be a large number of strategies available by varying hold times, and portfolio strategy as well as defining what transactions decisions are tied to what levels of the signal. Additionally, other signals from other sources might be integrated, both in selecting trades and also weighting portfolios. Clearly, the potential trading value of any signal will depend greatly on the trading strategy employed that uses it. Financial modeling expertise will be required to select the optimum trading strategy for any signal of interest.

We explored one news analytic approach for this in a blog posting. In that case we analyzed a trading strategy based on a change in sentiment in specific sources about a company.

According to the selected strategy, if positive sentiment was increasing over time we took or held a long position, while a decrease in positive sentiment led to taking or holding a short position. Evaluating the market performance of a paper portfolio based on these trading signals is displayed below:

This particular strategy did a good job responding to the market crisis in late 2008 but doesn’t fare well in less turbulent times. Perhaps this signal could be used in other trading strategies to improve performance.

Trading independent analysis

One may want to look for statistically significant relationships between two types of events, or events and continuous readouts that are not related to trading. In general, we have a collection of point processes and continuous data streams. Exploring if point processes are predictive of continuous processes can be done similar to the trading strategies discussed earlier. Examine the set of changes in the continuous variable following an event and determine if the behavior is typical or not.

For example, consider a set of momentum changes per day for a company when the event has not occurred. This collection of changes will have a mean and a variance. We also consider the momentum changes from the much smaller set of days following specific event occurrences.

We can use parametric (i.e. t-test) or non-parametric (i.e. wilcoxon test) approaches to establish the likelihood of the two sets of data having the same distribution. These approaches can establish a statistically significant relationship between the two signals, although not reaching the standard to determine causation.

Examining relationships between point processes can be performed in a number of ways. One simple approach is to look at the rate of occurrence in one of the event types in time periods before or after the other event type. Compare the observed rates in these time periods to the overall rates to assess the significance of the relationships.

10 financial modelling experiments for you to run with Recorded Future’s News Analytic Data
Here are some suggested analyses you can run with Recorded Future.

Can I build a profitable trading strategy using sentiment and momentum based metrics for different events types?
Can I detect times where sentiment/momentum for a company diverge from those for an industry?
Are certain events predictive of abnormal returns?
Can I define a set of Future occurring events that are predictive of market metrics like abnormal returns, volatility, or volume?
Can I incorporate Recorded Future news analytic content (events, or company metrics) in my existing models to improve predicted power.
Can I predict the times when my existing models fail.
Can I find collections of related events that are predictive of market metrics.
Can I assess the credibility of a source by looking at past predictions
Can I detect quiet periods for companies?
Are there differences between blog and mainstream sentiment and can I build a trading signal from this.

Getting Started with the News Analytic API

Users access the Recorded Future content via a web services based Application Programming Interface (API). Using an industry standard JSON format, many different languages and environments can be used to access the service including Python, Java, R, and Matlab.

We maintain documentation of our API as well as examples and have put together a tutorial showing how to use them. These samples are hosted on our new Google Code site, which is our central repository for hosting these such examples.

API user’s can download these examples and start accessing Recorded Future content immediately. Access to this documentation does not require an API license and anyone interested in a deeper and more technical investigation of the API and content can review the materials at the Google Code site.

Recorded Future Web Analytic Interface
Recorded Future also provides a user interface for interacting with our content. Quantitative traders may use this site to begin exploring the type of data we organize to look for potential signals and patterns that they can use to make trades on an ongoing basis.

The web user interface site can be used to support a hypothesis generation phase. Once hypotheses have been formulated they can be backtested via the API and if deemed valuable can be implemented as part of a trading strategy using future data obtained through the API.

Any pattern researched can be systemically monitored through the use of so called Futures, where a pattern is monitored and users notified via email if the pattern is matched – i.e. notify me as soon as there’s a product problem among pharmaceutical companies within a week of a product launch.

Conclusion

Recorded Future’s news analytic data contains discrete entities and events occurring in the past present and future as well as a (growing) number of derived continuous metrics generated from these events and entities. A web service API is available for investors to extract data sets of interest into their analytic environment of choice and historical data is available for building relevant models.

Once an investor has determined a useful model, realtime API queries can be performed to extract the latest data to be applied in the model. This suite of data and tools is currently in use by finance professionals and is available to others interested in adding news analytic strategies to their quantitative modeling approaches.

Live Webinar: Recorded Future for Quantitative Financial Research and Trading

2010-09-10T18:19:00.003-04:00

Join us on Wednesday, September 15, at 11am Eastern and learn how Recorded Future’s temporal news analytics content is used to support quantitative financial analysis and trading across asset classes including equities, fixed income, energy, etc.

Our CEO, Dr. Christopher Ahlberg, will introduce Recorded Future and its temporal analytics followed by an in-depth discussion of the Recorded Future data and web service capabilities led by our Chief Analytic Officer, Dr. Bill Ladd.

We'll demonstrate how, using computational linguistics, we extract events, entities, temporal cues, and metrics describing them from a wide variety of online media and index these temporally. This data identifies historical, current and expected future events as well as associated statistical measures such as momentum and sentiment.

We’ll show how our web service API integrates these events and measures into quantitative financial research and algorithmic trading strategies, supporting activities ranging from construction of alpha-generating signals to regime change detectors.

Register Here

When: Wednesday, September 15 at 11am Eastern
Where: WebEx (register for details)

We look forward to you joining us!

Market Cap Vs. Coverage

2010-08-15T17:52:00.008-04:00

You might think that larger market caps lead to more media coverage, and in general you would be correct, but today I’d like to drill down a bit into what this relationship really looks like. We first look at the simple plot of market cap vs. observations in our database (since April 1, 2010) for S&P500 companies:

In general, we see a general increasing trend but the variability is fairly high. In fact, the largest company, Exxon Mobil is ~35% larger than the next highest company, Apple, and yet we have observed 75% less content associated with it. In fact, 30+ companies have more associated content than Exxon Mobil.

Another way to look at this data is using log scales for each of the variables.

This certainly indicates a reasonably linear trend and indeed, the R2 for a regression line is ~.3. Looking at the industry sectors that seem to be present on the high side of the data for a given market cap, we see evidence of higher than average coverage associated with some of the companies in the Information Technology and Consumer Discretionary categories along with a few representatives from the Financial sector. In the lower portion of the data, we again see companies from Industrials, Materials, Staples and Financials.

When we drill into any local area of this dataset, we still see a high degree of variability. Consider the collection of companies between $10B and $20B in market cap (just over 20% of the total number of companies). We examine the distribution of the log of the market caps for these companies on a histogram and note that while the range of the company size is a factor of two, the range of the media coverage is roughly two and a half log steps, or roughly a factor of 300.

In fact if we look at the histogram for all of the companies, it does not look all that different:

So while there is a relationship between market cap and media coverage, it's not nearly as strong as we might think.

Taking a closer look at individual industries shows the relationships we observe on a per-industry basis. You can see different slopes, intercepts and R2 values for the different industries.

There are numerous implications to consider about this from a news analytic/trading strategy. Do the different levels of news reflect different levels of speculation about different companies? Are lower volume news companies more likely to respond to news events than high volume companies because of the scarcity of events?

In any case, this is a phenomenon one would want to better understand when using news analytics to support an investment strategy as companies of similar sizes can have very different media flow.

Predictions of "Futures"

2010-08-06T15:44:00.007-04:00

As I mentioned in an earlier post, one of our approaches to prediction is to gather, organize and present the predictions of others. Obviously, these aren't guarantees of future occurrences, but we feel there is value in having access to this type of information for a variety of reasons. Some of these are simply announcements of planned activities such as this news from March of future iPad availability:

http://www.9to5mac.com/ipad-april-3-pre-orders-march-25409682734

and some are mere speculations, such as when iPhones might be available on the Verizon network (January 2011?)

http://www.physorg.com/news197109182.html

In either case, if you are interested in a topic, you are probably particularly interested in associated likely future events.

When we harvest content, we capture both the publication date of the information and the event time of the events harvested. This event time can be in the past, present or future depending on the linguistic context of the content. As such, a certain percentage of events that we harvest are forward looking. The time spans for these future events can vary widely from a day to a month to a decade depending on the precision of the prediction.

For this analysis, I used our news analytics web service interface to collect our forward looking statements about S&P500 companies with a time span of a single day. Note that I'm not discussing events occurring after today (July 28, 2010) but rather events predicted to occur after the publication date. With these events in hand, it is relatively straightforward to assess what happened on the markets on these "future" days. I looked at "future events" predicted from the beginning of 2009 to the present. For each company with more than 5 predicted events, I looked at the local standardized trading volume for each day. This standardized the volume of the company on the previous 20 days and is essentially the volume for the company for the day minus the recent average volume for that company and then divided by the recent standard deviation of the volume for that company. I also tried using various other standardizations but the change had little impact in the overall analysis. I fitted a model across all of the companies looking at whether having an event predicted for a day predicted a larger volume than not. And on average, this turned out to be true. The model output from a model fit in R is below

Call:
lm(formula = LZvolume ~ future, data = datatable[datatable$futurecount > 4,])

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.045481 0.006486 7.013 2.36e-12 ***
future 0.403100 0.024010 16.789 <>
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 1.718 on 75694 degrees of freedom
(67 observations deleted due to missingness)
Multiple R-squared: 0.00371, Adjusted R-squared: 0.003697
F-statistic: 281.9 on 1 and 75694 DF, p-value: <>

The key observation here is that the future estimate is positive suggesting that on average the standardized volume increases around 0.4. Recall that with a standardized variable, 95% of the data is roughly between 2 and -2 (96% in our case) so a change in 0.4 is about 10% of the range. The R2 values here are relatively low suggesting that while on average the a future event is associated with an increase in volume, the overall data is quite noisy.

We then looked at whether the trend held up across different industries and found that generally it did. The estimates and P-values for several industry categories are listed below.

Industy	Future effect (change in standardized Volume)	P-Value
Industrials	0.51	3.03E-11
Health Care	0.64	4.73E-09
Consumer Discretionary	0.51	2e-16
Information Technology	0.33	1.23E-14
Utilities	0.32	0.1712
Financials	0.25	1.10E-06
Materials	0.85	4.10E-09
Consumer Staples	0.62	1.62E-10
Telecommunications Services	0.26	0.0173
Energy	0.43	0.0023

I took a more visual approach to this analysis as well. For each company, I split the volumes into two groups, those with an associated future prediction and those without. I used a Wilcoxon test (similar in concept to a t-test, but less sensitive to outliers) to compare these two groups for each company.

Using Spotfire to display the set of P-values I obtained (one for the statistical difference of the two groups for each company) gives the following histogram,

If there were no relationship between future events and our volume measure, we’d expect this bar chart to be relatively flat, with roughly 5% of the t-tests having a p-value less than 5%. In contrast, we see about 35% of our companies having significant differences between predicted event volatility and non-predicted volatility.

From our earlier analysis, the companies where future events are predictive of volume increases don’t appear to be segmented by Industry. So are they segmented by market cap, region, news coverage? That will need to be the subject of a future post.

Google Code Site and Examples

2010-08-02T18:43:00.002-04:00

We have just posted some examples of using our web services news analytics API from within Python and have put together a tutorial showing how to use them. It is hosted on our new Google Code site which will be our central repository for hosting these kinds of examples going forward. If you're a user of the API, make sure to bookmark it! We hope to fill it with examples of code you'll find useful for your own projects.

If you're interested in obtaining access to our API, please e-mail us at sales@recordedfuture.com.

Recorded Future and Quantitative Trading

2010-07-12T04:59:00.004-04:00

We have a new blog post on the conceptual ways that Recorded Future content can be integrated with Quant trading strategies. Take a look and let us know what you think: Recorded Future and Quantitative Trading

Predicting the Future with Recorded Future

2010-07-01T16:34:00.003-04:00

We often get asked about the quality of our predictions and our support for backtesting and I wanted to take a minute and put out something on this blog about that.

I think it helps to discuss two kinds of predictions you can do with Recorded Future. One thing that we do is aggregate and structure what others have said about the future and support the analysis of that information. Those aren't "our" predictions per se but rather our support for accessing and using the predictions of others. The second type are predictions an analyst specifically makes by using our content (perhaps in conjunction with other data). We have a couple of blog posts that highlight this approach that you can look over at
http://www.predictivesignals.com/2010/06/does-momentum-predict-higher-trading.html
http://www.predictivesignals.com/2010/06/news-sentiment-analysis.html
In these posts, we look at the relationship between content in our database and future events like volume or asset price. These involve postulating and testing specific predictive relationships.

So to consider backtesting, we record two timestamps for everything we capture, one is when the information was published and one is when the event is expected to occur (in the past, at publication time, or in the future)

Assessing prediction quality in either of these cases is currently best done using our web services analytic framework (JSON web service + Python/R etc)

For the first type of prediction, it is fairly straightforward to find content where event time is after publication time and to assess the reliability of those predictions. This is really assessing the reliability of the individual predictors we have captured in our system. In the second case where an analyst is exploring predictive relationships based on our content, it is easy to understand what was known when and to assess the quality of the predictive hypothesis. Again, the blog posts above are examples of those types of analysis.

Over time we will be adding additional support for backtesting to our product, both in the online and web service analytic frameworks.

News Sentiment Analysis

2010-06-30T10:27:00.006-04:00

News and Media Sentiment Analysis for Trading

The intersection of news, sentiment analysis and trading strategies has been considered for a number of years. Work ranging from Tetlock et al (2005) to the more recent efforts by Leinweber et al (slides) (paper) have established relationships between sentiment on publicly available news information and future changes in asset value.

We recently implemented a sentiment metric on entities (e.g. General Electric, Microsoft, etc.) and events (e.g. IPO Tesla, Product Launch Pfizer) that we process and decided to make an initial assessment of this metric in the context of trading strategies. Are there relationships between our sentiment metric and is there evidence that these can be predictive for asset prices. With these questions in mind, we executed a project to take a closer look.

We developed and tested our sentiment metric based on business related documents so it is generally tuned for company/financials related news. The sentiment metric is instantaneously measured on all entities and events in documents that we process and higher values of the metric correspond to higher levels of positive sentiment.

We examined the sentiment around a collection of earnings related content in the system and analyzed average sentiment measures around earnings calls. Earnings calls obviously have a large impact on asset values and one would expect that earnings calls surrounded by positive sentiment would be followed by a rise in price and calls surrounded by negative sentiment a fall.

Figure 1: Correspondence between calculated sentiment and the Close of the S&P500

First, we simply calculated our average sentiment measure over time for S&P500 companies compared to the close of the S&P 500. As you can see in the above graph (made using Spotfire Professional), our sentiment metric tracked the S&P fairly well although it does seem to lag the index for some of the time. This was actually very encouraging. We didn't include any pricing data in the development or testing of our sentiment metric so seeing this close a relationship suggested that we were measuring a real phenomenon with our metric.

Figure 2: Relationship between calculated sentiment on earnings call presentations and two selected companies

We next drilled down and took a look at a few individual companies and as you can see above, saw reasonable correlation with some companies such as Google and less with other companies such as Pfizer. Our final step in this first analysis was to take a deeper look at longer term phenomena. Specifically, we looked at the difference in sentiment scores between quarters and whether this value was useful in predicting asset values. We implemented a paper trading strategy to evaluate our sentiment signal based on earnings calls for S&P 500 companies back to 2006. We "bought" a stock the morning after the call if we saw an increase in the call related sentiment between quarters and "shorted" a stock accordingly if the sentiment fell. We then held on to the position until the next earnings call event for that company and then either maintained the position or reversed it based on the difference between the sentiment aggregated for the new quarter's call and the previous one. We looked at this on a return basis, rebalanced the portfolio daily to keep the positions in each company to an identical amount and averaged the return across the companies we were invested in at any specific point in time. In this approach, we ignored trading costs.

Figure 3: Returns from sentiment based trading strategy compared to S&P500 performance

The above plot illustrates the returns our approach yielded as well as the returns from the S&P500 over the three year period from April, 2007 to Mar, 2010. We started in 2007 because by this time we had sufficient historical content associated with the vast majority of the S&P500 companies to perform our calculations. Our total return for the this period was 15% (s.d. of daily return was 0.7%) compared to -18% (s.d. 1.9%) for the S&P500. We yielded 24/36 profitable months as opposed to the 18/36 profitable months observed with the S&P500. Our approach tends to under-perform the market when the market is rising, but over-perform the market, sometimes significantly, when the market is declining.

Taking a closer look at each of the three covered years independently confirms our view that the approach is generally doing well in market downturns, but is missing the upturns, suggesting that negative sentiment may be more telling in this context then positive sentiment.

Figure 4. Three 12 month segments of strategy performance vs. S&P500

Since we didn't account for trading costs and since the approach under-performs a rising market, we aren't done working with this signal. However, we were encouraged to see that the overall correlation we saw in Figure 1 can be used to derive a reasonable starting point for a trading strategy that outperforms the market overall for this three year period. Further investigation of this signal will include:

Subdividing by Industry/Market Cap - Leinweber illustrates that sentiment can have a different impact based on the market cap and industry of the set of assets evaluated. Perhaps this approach will work with technology companies like Google, but not Pharmaceutical companies like Pfizer.
Modifying position holding times - With essentially a three month holding period for each position, we are at risk for any other trends in the stock to change the fundamental behavior. Depending on when in the time the gains typically come with the stock, there may be room for optimizing with shorter holding times
Incorporating stop loss and profit taking rules - Another strategy to reduce holding times is to base them on stock performance criteria
Improved use of the sentiment scores. Currently we are only examining the change in sentiment direction, but not considering the magnitude of that change or comparing the magnitudes across companies. Restricting to the larger changes, either in relative or absolute terms may improve performance.
Incorporating other content. We calculate other metrics such as momentum and a hedging score. Using these scores could help.
Only include documents/events that are decisively not about price change that has already happened (e.g. avoid documents saying "Pfizer was down 3% in late trading yesterday").

Our efforts here support the current research in the area, that there is promise of finding alpha from the linguistic analysis of publicly available content. The challenge is to find and implement the optimal approaches.

News as circuit-breaker

2010-06-30T10:22:00.003-04:00

While researching another project, I ran into these two posts discussing news analytics.

http://ftalphaville.ft.com/blog/2010/01/26/134561/rise-of-the-news-reading-machines/

http://ftalphaville.ft.com/blog/2009/11/04/81356/why-trading-machines-dont-like-news-releases/

One quote that particularly caught my attention was this

"If something big comes out on a company, computer traders can set up a sort of circuit-breaker which trips and pauses the computer trading programme."

This echoes something we've been hearing a lot from the people we speak to. A first approach to taking advantage of news and pre-news (government filings, twitter, etc) may be in using these signals to alert a trader that their trading strategies might be inappropriate for a given time period. Perhaps if you know that a major product release, government approval or court case is being announced on a specific day, its not a day you trade that stock with your established trading strategies. At least until after the initial activity around the event.

Since assessing future events is one of the things we do, we'll be taking a closer look at this in a future blog post. If you'd like to learn more about our media analytics check us out over at Recorded Future.

Does Momentum Predict Higher Trading Volume?

2010-06-08T10:41:00.007-04:00

[Originally posted at the Recorded Future blog]

Every day, billions of dollars change hands in the U.S. stock market. In a single trading day, a big company like Exxon Mobil might see $5 billion of its shares change hands. This quantity might not mean anything to an individual investor looking to buy or sell a few shares on Etrade, but it means a lot to a large institutional investor looking to unwind a $500 million position in a particular stock.

With a position that large in a given stock, the aforementioned investor is unlikely to be able to trade completely into or out of their desired position the stock without adversely affecting its price over the course of a single day. Instead, they are likely to work the trade over the course of several days. But just how much of their position should they trade in a single day? The answer to this question is in part dependent on that investor's expectations of how much will be traded in the greater market.

So what are the factors that drive expected trading volume? In part, the previous day's, and perhaps month's trading volume impact expected near term trading volume. Additionally, a significant piece of news about a company, such as a product release or earnings announcement may drive trading volume up as traders more actively move shares around and the market determines the “right” price for a stock.

News/media flow (remember it's not only about news in it's classic sense - it involves everything from regulatory filings to blogs) is typically difficult to quantify. What constitutes news? How much of it is out there? Is it actually relevant to a company's value, or is it just PR fluff? Additionally, obtaining historical data about the news and using it in a statistical model is often quite difficult.

Using data from Recorded Future's advanced platform for processing the semantic structure of the web, I have taken a simple autoregressive model to predict trading volume, and augmented it with quantified information about the news about companies in the S&P 500 Index, and will demonstrate that incorporating this information into the model has a statistically significant effect and may provide more accurate predictions about future trading volume.

Experimental Setup and Data

I obtained my information about individual stocks from August 1, 2009 to April 10, 2010 from Yahoo finance with the help of the Rmetrics software package. Yahoo's historical quotes provide pricing and share volume information for companies on each date. I estimate daily dollar volume for a company as its share volume * closing price, because Volume Weighted Average Price was not publically available. I made use of the timeSeries class on this data to calculate lagging and trailing moving average numbers for each stock on each day.

Using the Recorded Future API together with the R language, I was able to pull aggregate news information about all of the companies in the S&P 500 over the same period. Included in this information was each company's “Momentum” on any particular day. Momentum can be thought of an aggregate indicator of news or “buzz” behind a company on a given day.

Using the R software, I then combined this data to derive a time-series cross-sectional dataset, representing dollar volume and news sentiment for all companies in the S&P 500 over the time period.

Model Specification

I propose a simple autoregressive model for predicting trading volume with a simple moving average term as follows:

DVt = a*DV(t-1) + b*SMA(DV, t-1, t-20) + et

Where DVx is Dollar Volume at time x, SMA provides a simple moving average function on a range of time periods, and et is the error term at time t.

I then augment this model with a momentum term. Because we are estimating Dollar Volume, and that raw number is highly variable (and largely dependent on Equity Market Capitalization of a given firm), I scale momentum by the moving average term.

DVt = a*DV(t-1) + b*SMA(DV, t-1, t-20) + c*(MOt-1*SMA(DV, t-1, t-20)) + et

Experimental Results

I constructed two models using the R "lm" function:

> dflm <- lm(Dollarvol ~ 0 + lDollarvol + smaDvol.Dollarvol, seriesdf)
> summary(dflm)

Call:
lm(formula = Dollarvol ~ 0 + lDollarvol + smaDvol.Dollarvol,
data = seriesdf)

Residuals:
Min 1Q Median 3Q Max
-5.060e+09 -2.277e+07 -2.686e+06 1.755e+07 1.597e+10

Coefficients:
Estimate Std. Error t value Pr(>|t|)
lDollarvol 0.513351 0.003237 158.6 <2e-16 ***
smaDvol.Dollarvol 0.477892 0.003600 132.7 <2e-16 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 1.71e+08 on 72110 degrees of freedom
Multiple R-squared: 0.8539, Adjusted R-squared: 0.8539
F-statistic: 2.107e+05 on 2 and 72110 DF, p-value: < 2.2e-16

> dflmMo <- lm(Dollarvol ~ 0 + lDollarvol + smaDvol.Dollarvol + smaxlMo, seriesdf)
> summary(dflmMo)

Call:
lm(formula = Dollarvol.1 ~ 0 + lDollarvol.1 + smaDvol.Dollarvol.1 +
smaxlMo, data = seriesdf)

Residuals:
Min 1Q Median 3Q Max
-5.039e+09 -2.215e+07 -2.284e+06 1.813e+07 1.597e+10

Coefficients:
Estimate Std. Error t value Pr(>|t|)
lDollarvol.1 0.513193 0.003237 158.54 < 2e-16 ***
smaDvol.Dollarvol.1 0.471645 0.003817 123.56 < 2e-16 ***
smaxlMo 0.077162 0.015683 4.92 8.67e-07 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 170900000 on 72109 degrees of freedom
Multiple R-squared: 0.8539, Adjusted R-squared: 0.8539
F-statistic: 1.405e+05 on 3 and 72109 DF, p-value: < 2.2e-16

We can see that the addition of the Momentum term provides a statistically significant enhancement to the estimate at the 0.001 significance level as well as at the adjusted R-squared level.

Drawbacks

Note that both of these models are linear, when in fact the relationship between trailing momentum and current momentum may be non-linear. For example, based on the roughly lognormal distribution of Dollar Volume, a log-linear model may be more appropriate. Further, the error term may not be normally distributed or exhibit heteroskedasticity, invalidating some model assumptions and providing weaker than expected predictions.

Additionally, I have built these models on less than 1 year of economic data, and not taken into account other factors that may affect dollar volume. These include, but are not limited to - exchange/OTC market effects, non-linear market capitalization effects, industry effects, seasonal effects, and the effects of stocks with multiple share classes (e.g. Berkshire Hathaway). The relatively short time span does not capture a full picture of trading volume over the course of the greater economic cycle.

What's Next?

Whereas this is an interesting result in itself there are many types of deeper analysis to be done. What about volatility and price? What about breaking down Momentum from news/media flow by type - mainstream media vs. blogs vs. government filings, etc. ? What about exploring the effects of news/media Momentum to other asset classes?

If you'd like to try this yourself, contact us to gain access to our API!