Publications

by Vincent Granville, Ph.D. / CEO of Data Shaping Solutions
(www.datashaping.com)

In this article, we investigate three potential problems related to
using and designing statistical trading strategies:

1. Too many parameters
2. Selecting a metric to measure return
3. How to efficiently use historical data

Over-parametrization can best be described as using a trading model
with more than four parameters. The parameters are usually optimized
on historical data.  It is well known that this approach yields
catastrophic results.  However, it is possible to successfully work
with 6 parameters if you follow a few rules.  First, test the
strategies on at least 200 trades. Then keep only the reliable
strategies.  To achieve this goal, test billions of parameter
sets. For each set, introduce various forms of noise to see how it
impacts performance.  Then discard the vast majority of parameter sets
that are too sensitive to noise. Do NOT look for the most efficient
strategies, but instead for a strategy that is high-performing AND
noise-resistant. To further improve the results, frequently update the
parameters.

When designing or using a strategy, one wants to measure its
effectiveness.  Always use two metrics: one that measures the average
return over a long enough period of time (corresponding to at least
200 trades), and one that measures the volatility of the
strategy. Strategy volatility is different from stock price
volatility. While volatile stocks are interesting for short-term
traders, volatile strategies are not. To reduce volatility, discard
strategies with a large time span of negative growth at any give
point.

Now, with the right metrics and the right number of parameters, let's
try various kinds of testing. We will focus here on a strategy with
daily buy and sell signals. One obvious way to simulate actual trading
is to fit the strategy parameters with data that is between 31 and 180
days old, and then check the actual performance on the last 30
days. This is an improvement over classical backtesting. Also, when
fine-tuning a strategy, put emphasis on performance during the last
few weeks of the test-period. It helps to work with stocks that
exhibit a rich variety of price patterns. Technical notes on all those
topics are regularly published in Data Shaping's newsletter.

Contact:

Vincent Granville, Ph.D.
vincentg@datashaping.com
www.datashaping.com