How to Build an Adaptive Trading Strategy in 3 (Not-So-Easy) Steps

31Aug08

I’ve received a lot of reader feedback lately on my recent posts showing that the fundamental characteristics of the markets are evolving (examples) and the benefit of building trading strategies that are adaptive (such as our YK Strategy).

In this post, I want to talk about how to build an adaptive trading strategy.

My definition of an adaptive strategy is one that doesn’t have hard and fast rules, but rather, “learns” from the markets. Each day, an adaptive strategy scraps all of its assumptions and reanalyzes the markets based on the newest available data. In theory, such a strategy should be able to adjust to not just changing market conditions, but changes in the fundamental way the markets work.

Now, I’m not going to say exactly how we built our YK Strategy (this is after all a zero-sum game we play), but I do want to walk through the thought process we used. I’ll frame most of the process in a series of questions.  I think it’s important to note that there is no right or wrong answer to these questions – our solution is our solution, but by no means do I think it’s the only solution.

This post will be long, but the basic flow will go like this: (1) what do we want our strategy to study, (2) how do we want our strategy to assess what it studies, and (3) how do we want our strategy to learn from what it assesses? Here we go…

WHAT DO WE WANT OUR STRATEGY TO STUDY?

Every adaptive trading strategy has to have some bounds on what data it’s aware of. Regardless of what those fancy neural-network types tell you, it cannot (for the foreseeable future) be a human brain. So we have to tell our strategy what it might consider important.

Is it relative strength, moving average crossovers, company earnings, or the price of tea in China? That’s up to you to decide. But there are two competing schools of thought.

The first says let the strategy study as much data as possible – throw millions of bits of data in there and see what comes out the other end. The other says give the strategy a small group of data that has proven historically effective in predicting the market (even if what it has said about the market has changed).

The first school is good because it captures everything that might move the market but bad because it’s much more susceptible to curve-fitting and very computing-intensive (and vice-versa).

The best answer? Up to you.

HOW DO WE WANT OUR STRATEGY TO ASSESS WHAT IT STUDIES?

So now our strategy has a set of data it’s looking at and it knows how the market has performed over some given period when combinations of that data existed, but how does our strategy assess that performance?

It could look at average return, risk-adjusted return, do something like a multivariate regression, or use any other number of other obfuscated ways. But regardless of the method chosen, the strategy has to have a means to account for a couple of result-skewing factors: fat-tail days and the broader trend.

Fat-tail days (think October, 1987) have a huge impact on results even though they could have been caused by factors completely unrelated to what the strategy is assessing. The strategy must have a way to wrangle the influence of these big up/down days.

The broader trend can change how the market performs under different data combinations – the strategy must know that just because something worked in the bull market of the 1990’s, it’s not necessarily going to work in the bear of the early 2000’s. The broader trend must either by predicted by the model or the broader trend must be discounted to create a trend-neutral prediction.

The best answer? Up to you.

HOW DO WE WANT OUR STRATEGY TO LEARN FROM WHAT IT ASSESSES?

This is where the magic happens.

So now the strategy has a basket of data it’s aware of and it knows how to assess the market’s performance under different combinations of that basket of data, but how does it learn from those assessments day by day?

The possibilities are endless and the answer is open ended, but there are a couple of important factors that must be considered: is it a light switch or a dimmer and how quickly to learn and forget.

First, is it a light switch or a dimmer? Do we want our strategy to move from always seeing some combination of data as bullish or bearish like a light switch that’s always either on or off, or do we want it to move between different levels of bullishness or bearishness like a dimmer with a whole range of on or off. In the case of the YK Strategy we’ve taken the dimmer approach – each day is anywhere between 100% bearish and 100% bullish (and every step in between) depending on the strategy’s confidence in that day’s analysis.

And second, how quickly do we want our strategy to learn and forget? Do we want our strategy to treat every new observation with equal weight or do we want to overweight more recent observations? Do we include the entire market’s history or only the last X number of days or years? What happens to older data – does it linger on in the strategy’s memory or is it completely forgotten?

The best answer? Up to you.

CLOSING COMMENTS

There it is. I hope I haven’t been overly vague or confusing. This is a very difficult topic to discuss in just a single blog posting and perhaps I’ll do some follow up posts in the future to break down different parts of this one with concrete examples.

More importantly though, I hope that I’ve gotten some creative juices flowing for other traders out there to begin thinking about applying these adaptive concepts to their trading.

Happy Trading,
ms

 

To stay up to date with what’s happening at the MarketSci Blog, we recommend subscribing to our RSS Feed or Email Feed.

Filed under: Evolving Markets, Trading Strategies   |  12 Comments