Sunday, March 23, 2014

The Psychology of Quantitative Analysis

Early this morning I began my weekly routine of model building.

As a thought experiment, imagine taking every technical indicator out there and conducting a big factor analysis.  The factor analysis would reduce the number of indicators to a smaller cluster of factors that are relatively uncorrelated.  

This is important because it turns out that many indicators, from a purely mathematical vantage point, are measuring the same thing.  A 14-day RSI, for instance, may correlate very highly with a 14-day rate of change and a 14-day stochastics.  If you look at all three indicators, you're really looking at one variable measured three ways, not three unique variables.

What you really want are unique variables that are significantly correlated with forward price movement.

The bad news is that the many technical indicators out there really just boil down to a handful of unique variables.  The good news is that, overall, these unique variables do possess statistically significant predictive validity with respect to the prospective movement of stock index prices.  The challenging news is that even this significant predictive value leaves the lion's share of the future movement of stock index prices unexplained.

So imagine I identify a handful of unique predictive variables from among the large array of technical indicators and I identify the expressions of those variables that minimize their overlap.  From these few variables--it's important to reduce the likelihood of overfitting the data--I conduct a regression analysis and arrive at a statistically predictive model over an identified market regime.

Over the regime, let's say the model has been 65% accurate in forecasting the direction of S&P 500 Index prices over the next three trading sessions.  When the model has given its strongest signals (top quartile of forecasts), the average three-day gain in SPY has been .64%.  When the model has given its weakest signals (bottom quartile of forecasts), the average three-day loss in SPY has been -.28%.  This performance has held up well in out-of-sample testing.

Is this a good model?  It possesses a statistically significant "edge" and yet its R-squared, the amount of variance accounted for in future index prices, leaves about 90% of future action unpredicted.  A full 35% of the time, the model has been wrong in identifying future price direction.  And yet, a model that gets market direction right two-thirds of the time is better than throwing darts, assuming that we remain in the stationary regime that we backtested (an important assumption).

What quantitative work accomplishes for me psychologically is that it clearly identifies what is known and what is unknown.  It gives me a sense for when there is an objective edge and it provides a sense for the fragility of that edge.

Does quant modeling "take emotion out of trading"?  No, but it does something more important.  It replaces the emotions associated with overconfidence and confirmation biases with a different set of emotions: the humble respect for what is unknown, the desire to expand the frontier of the known, and the felt imperative to quickly adapt to what Victor Niederhoffer calls "ever-changing market cycles". 

Further Reading:   
Predictability as a Market Variable

Quant Reading:
See publications section of Marcos Lopez de Prado's site