Sunday, July 17, 2016

A Systematic Approach to Discretionary Trading

There's a lot to be said for wandering.  Pick a promising area, wander about, and you can discover quite a bit.  Discovery is all about open eyes and open minds.  

Here's how I've been wandering through markets lately:

Speaking with trading and investment professionals, particularly in the quant world, I've been struck by the fact that the models they employ to guide their decisions are not the kind of models we typically read about in trading texts.  There are no technical indicators or chart patterns in their inputs.  Nor are there any inputs pertaining to company earnings, economic growth, upcoming central bank meetings, or geopolitical events.  Rather, returns from markets are broken down into basic "factors", such as momentum/trend, volatility, value, and carry, with models designed to capture these factor-generated returns.  

A discretionary trader might justify trading a trend following method or a mean-reversion/reversal method based on his or her "personality".  This is nonsensical to the money managers I speak with.  It's like saying that I'm right handed, so I'll only take right turns in my car.  If returns come from a variety of factors, the best performance can be achieved by trading signals derived from each of these factors.  This will diversify returns and produce better risk-adjusted results.

A comparison of returns from quant asset managers vs. discretionary trading firms finds that returns indeed have been better from the former group.  That is not simply because these funds are quantitative--I can think of quant funds that have lost money lately.  Instead, the superior performance comes from generating returns from multiple factors across multiple time frames.  Many good models producing independent, positive returns--not necessarily eye-popping ones--can combine to form a robust P/L stream.

So I began my wandering.  Of all the market data sets I track, I identified the ones that: a) produced the most reliable and valid trading signals; and b) had very low correlation to one another.  A total of six variables popped up.  To my surprise, one was based upon volatility; two were derived from cycle-based forecasting methods; two were based on momentum (trend); and one was based upon value (mean-reversion).  I built six simple forecasting models based on the six variables and then combined the model outputs into a single "committee of experts" signal.  (See this article for an overview of creating ensembles of forecasting models).

I was surprised by the degree to which the trading signal from combining the individual forecasting models handily beat any of the individual models.  Still more surprising from my perspective was that the combined trading model seemed to "know" when to trade like a trend trader, when to trade like a reversal trader, and when to make money from shifts in volatility.  The model seemed to work by navigating the ebb and flow of factors.   

Most surprising of all, however, was that when I trained the models to forecast shorter-term price change, certain forecasting models dropped out, some received extra weighting, and some less.  This raised the possibility of using shorter-term models in a Bayesian fashion to navigate longer-term signals.  That is, you would use short-term forecasts and returns to change your assessment of the odds of a longer-term forecasted move playing out.

I'm not at all convinced that this means we should all toss our experience aside and become programmers, statisticians, and systematic traders.  Could it be that the discretionary "tape-reading" skill of the short-term trader can help navigate shorter-term forecasts, just as those shorter-term forecasts can help us participate in longer-term forecasts?  In other words, such trading would be neither wholly discretionary nor wholly systematic.  It would be discretion--with road maps.

But isn't that the way we travel when we drive cross country?  We don't simply rely on feel and intuition; we look at maps and we rely on GPS signals.  On the short time frame, however, we *do* use our feel to navigate lane changes, select optimal places to stop and rest, and adjust our speeds to road conditions.  The skilled driver has experience and road feel--and maps and GPS signals from trusted data sources.  Perhaps the skilled trader is not so different, navigating moment to moment price action, even while benefiting from the road maps of forecasting models.

Further Reading:  Factors and Short-Term Market Returns