Back in the olden days we used Javelin software to manage and distribute a large data base of rail carloading data to sellside analysts. Railshare is still available at railfax.transmatch.com. Behind the graphs and reports we create on the internet and for our clients (mostly sellside analysts) still sits a copy of the original Javelin software. It did (does) a great job with time series data. Cleverly we decided to plug in commercial data bases and see what happens. Over the years we designed nifty valuation models for Alcar, GE, Verizon, Soros and a host of Japanese brokers using stock price data and fundamental numbers from Compustat. Nomura came to us and asked if we could create a program trading model based on technical market signals used by Japanese investors. We did. They were happy. So we backtested the numbers with stunning results. Nothing worked in the backtest. We had a end of day model that didn't allow us to hide poor performance.
What to do? Well the finance textbooks of the time said the only way to outperform the market was to take systematic risk. Look up arbitrage pricing theory on Wikipedia. Sounded logical and we did have APT numbers (factors) from Alcar and Tony Elavia. So we plugged them into our Nomura model of the SP500 and found voila...something. Depending on which factor was dominant (operative) we found autocorrelation. Sometimes the high factor stocks outperformed the low factors. Then they reversed. The delta between the highs and lows trended. When the delta turned it usually turned enough to prompt a reliable signal.
The factor loadings we got from Alcar and Voyageur worked but not that well. As they were largely constructed with years of historical data they often gave mushy signals. That was ok for their clients who were only trying to immunize their portfolios from risk. But we wanted to maximize risk..to isolate the portable alpha from systematic risk. More later.