Sunday, August 09, 2015

Forecasting Next Day Price Change in SPY

Every so often, I'll post on topics relevant to my own trading, in hopes of inspiring fresh directions for active traders.  This weekend I undertook an exercise to develop a predictive model for the next day's returns in SPY.

Above are the results for the simple model.  The data are divided into quartiles, with the top quartile representing the most bullish predictions and the bottom quartile representing the most bearish predictions.  The model consists of only three variables, which I would describe as:  1) the amount of buying and selling pressure for the previous trading session; 2) the number of stocks in the SPX universe showing price weakness over a multiday period; and 3) the number of stocks in the SPX universe showing price strength over a multiday period.  The model lookback period was January, 2014 - present.

Here are a few takeaways from the exercise:

1)  As in other models I've built, the predictive value tends to be at the extremes of the distributions of the predictor variables.  This is a way of saying that the relationship between predictors and forward price change is not a simple linear one.  Being able to model those nonlinearities results in superior price forecasts.

2)  The model is statistically significant, but accounts for a modest proportion of overall variance in next day outcomes.  The model provides an edge, but significant uncertainty and randomness are present.  Even the best models require sound money/risk management.

3)  None of the three predictor variables is a traditional measure from technical analysis.  When I add traditional technical indicators to a stepwise regression process, the technical measures are generally not significant predictors in their own right and never add predictive value to the three model variables.  Parsing market data in unique ways promises unique forecasting power.  Measures most commonly followed by traders possess little forecasting power.

4)  Lookback period matters.  The model that is robust in 2015 does not perform well on 2008 data.  Before searching for predictive relationships, it's necessary to identify a stable lookback period that captures strong, flat, and weak markets.  All models assume that the stable regime of the recent past will persist into the immediate future--an assumption that will be incorrect at times.

5)  The hit rate for the top and bottom quartiles is around 60/40.  Again, it's an edge, but it leaves plenty of room for error.  The hit rate for the middle quartiles is closer to 50/50.  

6)  I have other models that forecast SPY price change over 3-5 day horizons.  There is value in lining up the daily forecast with the swing forecasts.

For me, the sweet spot of trading is seeing those occasions in which the action of the tape lines up with the anticipation of the market forecast.  The tape comes first:  I want to see expanding buying interest in order to go long and selling interest if I'm to be a seller.  When the forecast lines up with the tape, those tend to be occasions when it's useful to bump up the risk taking and extend holding periods for potential trend days.  An additional value of these models is that they keep me out of bad trades: I do not fade the market when forecasts are in the top and bottom quadrants.

For those who might be interested, Monday's forecast is the first positive one in several sessions, with a forecasted gain of +.11%.  That is in the second quartile; similar readings have been up 17 times, down 12 times for an average gain of +.12%.  My swing forecasts are neutral.  In all, Monday is not a day with a strong edge from the models.

Further Reading:  Integrating What We Know and What We Feel as Traders