Hack the market

The behaviour of the two strategies we described last time are easy to describe intuitively and they are furthermore in-line with what one might expect from mean-reverting and trend-following strategies.

A mean reverting strategy will tend to make many, reasonably frequent, constrained-size successful trades and then suffer a big loser.

A trend-following strategy will tend to suffer many, reasonably frequent, constrained-size losses and then celebrate a home run.

The two strategies we described last time have these characteristics and the only constraint we’ve put on them wrt to their big loser/winner event is that they’re both deployed with day-trading schedulers so that they start trading at a fixed time in the morning and flatten themselves at some fixed time later in the day. The same basic characteristics would emerge for other such fixed periods of time as well.

The first strategy we’ll consider is the mean-reverter. I ran the strategy across the most-active crude futures contract since the beginning of the year, permuted 100 different ways. Since this is a relatively “close-to-the-market” strategy which requires simulation of limit orders, I used full tick data including the order book so I could best simulate executions. The results are about as we expect – the strategy makes a lot of small wins and then suffers a big wallop. The most profitable instance of the strategy had the following distribution of returns:

His wins clustered around the expected return indicated by the minimum spread parameter – in this case a bit more than $500. His losses were much more broadly spread and his worst losses were substantially greater than his best wins. His win ratio was right about 70%.
Looking across all of the permutations of the reverter strategy, we see the same theme more fully fleshed-out:

To the right of zero (profitability), we see a pretty dense collection of returns. Indeed, depending on what setting we’ve specified for our minimum spread parameter, we see a further clustering around that expected return. And again we see that the worst losses far exceed the best wins. In aggregate, the strategies had profitable trades nearly 75% of the time, but their overall profitability was negative as the big losses swamped the frequent wins.

The trend-follower provides essentially the mirror-image. In aggregate, the distributions were:

We see that now our long tail extends in a favourable direction and we still see the clustering of many small – now – losses. On aggregate, these strategies traded successfully only about 23% of the time, but when they won they sometimes won big.

Although the trend followers did better on this particular series of data, it’s important not to extrapolate from this test that it’s fundamentally a better strategy. More likely this skew in our results is one of simple luck or a characteristic one might expect from data which hasn’t been de-trended.
Either way, the important thing to note about the results we see before us isn’t the particulars of which strategy did better in these specific tests, but that the two strategies showed fundamentally different behaviours. Neither strategy, as presented, looks particularly good to me! But the fact that we can see how each strategy has an underlying character that we can engineer and then tune for differing market circumstances is critical.

Next time, we’ll talk about how this result should color our practice of strategy development and our perspective on performance analysis.