Investing

Most professional traders cannot reliably beat passive market index returns. Specifically, over the course of 10 years, 71% of professionally managed funds performed worse than what would have happened if one simply followed the publicly available S&P 500 Index (Burton G. Malkiel).

The remaining 29% that were able to beat the market are either lucky, unfairly using asymmetric information, or have achieved truly differential insight. We believe the latter is possible, and have built an algorithm to analyze free text, ideas, and opinions within news to predict future price patterns based on our expected social reaction to news items.

We have been successful in outperforming the market in both extensive backtesting and forward-testing, and are now seeking investment capital. We invite you to learn more about our algorithm, and contact us with your investment interests.

Theory

If the majority of professional traders cannot beat the passive market, then why would anyone pursue active portfolio management? We believe tht the answer mainly lies in the false sense of security gained from technical (or chart) analysis. Although Economists have noted that stock prices move at random with no predictable patterns given past pricing, technical analysis is still widely used.

Why? For every 1,000 backtests that a chartist performs on predicting future prices from the past, at least 1 will emerge as "profitable" - finding a set of coincidental past price patterns that have been reliable, but have no bearing on the future. Assuming that prices are not able to predict future prices alone, then half of the chartists should expect to beat the market given pure luck, while the other half loses. Over the course of 10 years, the failures will drop out of the game as new entrants participate in this coin-flipping contest, naturally leaving a minority set of survivors that were able to beat the market - the 29% (Burton G. Malkiel).

Instead of relying on past prices to predict the future, we instead rely on news. The Efficient Market Hypothesis suggests that price movements are inherently nature because they follow reactions to unexpected news, which is also inherently random. We have adopted this methodology to develop our technology which solely looks at news to outperform normal market returns.

Technology

To fully understand how the public comprehends and reacts to the news is a daunting task, and we believe that we have only scratched the surface. We use a rich combination of natural language processing, machine learning, and knowledge engineering techniques to model the state of the world as interpreted from news outlets that we listen to.

Our system has two parts: signal generation and trading interpretation. For every stock on every day (trading or not), we generate a signal (from negative to positive infinity) that we use as a prediction of the intraday change for a stock, had there been trading on the day. These signals are generated a day in advance and are learned strictly from prior news and price movements. Although we do look at past pricing to train our algorithm, we do not use the past prices to predict future prices, as does technical analysis.

After generating signals for each stock on a daily basis, we purely analyze the signal patterns in forming the number of shares we would like to buy or sell of the stock. To simulate real-world conditions, we give our algorithm a fixed amount of cash and let the algorithm possess a long-only portfolio, purchasing stocks when cash is available, and selling stocks that are in the portfolio. We began by backtesting our algorithm with a fixed amount of cash on January 1, 2000, and then let the algorithm run live on July 11, 2011 when we were happy with our results.

Backtesting

Using historical pricing and news, we were able to gather enough data back to January 1, 2000, spanning over 11 years of backtesting data for us to observe our algorithm's performance had it been running live in the past. It was important to use such a long timespan such that the algorithm could be simulated over a generic enough set of events such that the system is not over-fitted to a specific time window, such as an election year or financial disaster in 2008. Backtesting occurs by feeding the algorithm prices and news as they become available over the simulation; the algorithm is in no way allowed to view what happens after the simulated date that it's on. We conducted this backtesting up to July 8, 2011.

Before evaluating what would be good performance for our algorithm, we first needed to establish a baseline - what would have happened if one held one share of each of the current S&P 500 stocks back in 2000 (purchasing newer stocks as they came available). We observed an average annual percentage yield of 8.8643% with a standard deviation of 18.8886%. We set our goal to attain a higher annual return than 8.8643% with the strict condition that we may not accept a standard deviation above 18.8886%, as this would be indicative of an over-fitted model that made its fortune by chance in some isolated time window.

How did our backtesting go? We were able to lower the standard deviation to 16.1947%, implying that our algorithm exhibited less risk in its portfolio selection over the baseline. What about percentage yield, did this also slightly improve. No. It improved a lot - all the way to 16.8341%, an average annualized market outperformance of 7.9698% with less risk. Be wary of competing services that promise unreal APYs, as these are indicative of over-fitted models with high variances that most likely profited over isolated events.

With these results, we were highly confident that our algorithm would be able to outperform the current S&P 500 and have been running the algorithm live since July 11, 2011, advertising our performance on our homepage. So far, our live performance has matched our backtesting expectations, mainly because we achieved a lower standard deviation than what the passive index did throughout backtesting.

Performance

Because we backtested on news, rather than prices, we believe that our backtests were not subject to the normal bias and over-fitting that most technical analysis techniques face. There was only one way to know for sure, and that was to run live - let the algorithm make its desired trades before the market opens every morning, and then evaluate these trades the following day when prices become available. We have been doing this for over 100 days and our live results are consistent with our backtested results: we achieve a higher return and lower standard deviation than does the market.

Contact

Please contact us if you are interested in investing in our algorithm. Although we are interested in anyone willing to invest a substantial amount ($100,000+), we prefer investors who can meet either of the following criteria, or have prior experience in portfolio management:

Have an individual annual income greater than or equal to $200,000 per year, or joint annual income of $300,000 or more over the previous two years.
Hold a net worth of $1,000,000 or greater either individually or jointly.

Our fee structure is to be decided, but we will only ask for a commission on the profits above the S&P 500 Index; we are not asking for a fixed management fee. Please contact us with your interest level as soon as possible - we will not accept further investment money once we reach a given threshold.