Monday, June 25, 2007

Trading by Regimes

A regime is a relationship that exists in recent history between one or more variables and prospective price changes in a trading instrument.

Some regimes describe intermarket relationships. For instance, we recently saw a period in which falling bond prices were associated with falling stock prices.

Other regimes describe relationships between market indicators and price change. For example, we could see how prices behave following instances in which we hit a threshold number of stocks making new price highs or lows.

Regimes do *not* attempt to describe relationships that will be present forever, or even for long periods of time. For that reason, regimes are not meant to be mechanical trading systems. Regimes come and go; that is their nature.

The proper way to think of regimes is as "the rules that the stock market has been playing by in recent market history". Recent history could be anything from the past several years to the past several days. Regimes identify the variables that have been associated with directional price movement during market periods that are similar to today's market.

When trading a regime, you make the base assumption that today's (upcoming) market will behave like the markets from recent history unless:

1) There are fundamental, market-breaking news items or economic reports moving markets sharply;
2) Volume and volatility today are meaningfully different from the recent past;
3) Interest rates, currencies, oil, etc. are behaving abnormally relative to their recent past.

In other words, you assume the very near-term future will look like the recent past *unless* there are distinctive indications to the contrary.

Regimes provide trading ideas; they are a heads up that tell you to take note: The market has behaved this way in the recent past. Sometimes regimes keep you out of a bad trade idea; sometimes they confirm ideas you develop from other sources.

Looking back over the past two weeks of trading (N = 4034 trading minutes), I found that, when the raw 60-minute average NYSE TICK was -100 or less (N = 210), the next two hours in the ES futures averaged a gain of .27% (204 up, 6 down).

At the other end of the spectrum, when the 60-minute TICK averaged +500 or more (N = 252), the next two hours in the ES futures averaged a gain of .33% (164 up, 88 down).

Across all other occasions over the last 10 days (N = 3572), the ES futures averaged a two-hour loss of -.01% (1657 up, 1915 down).

In other words, we've seen a bullish edge when the TICK has been very weak (reversal effect) and also when it's been very strong (momentum effect). Everything in between has led to subnormal returns.

There are many other regimes at play involving the TICK and other market indicators. When you see different regimes pointing to similar conclusions (reversals of weakness; continuation of broad strength), that becomes a useful trading concept for the coming session. Such a concept was particularly helpful to me in this morning's trade.

Markets play by rules. The rules change. The key is figuring out when fundamental shifts in markets are creating rule changes and when regimes will persist at least one more day.


How to F*** Up a Trade Setup

Oil Prices and Stocks

Historical Patterns as a Heads Up in Trading


TJ said...

I'm confused by this -- isn't finding short-term correlations without any kind of fundamental causation analysis essentially just data mining?

To determine that this isn't just a coincidence, shouldn't you do something like create "regime change finder" methodology and back-test it?

Brett Steenbarger, Ph.D. said...

Hi TJ,

Great question. One aside, however: I am a strong proponent of data mining. There is a huge difference between proper data mining and the overfitting of data. See for quite an education about data mining and its applications.

That having been said, I do think there's a huge risk of finding regimes in random correlations. I'm sure there's some pattern of baseball scores somewhere that correlate with recent stock market results. That's not to say we should trade on those data!

For that reason, it's imperative to select candidate predictors for regimes that are known to be associated with market changes over different cycles and that have a logical connection to stock market behavior.

Your idea of fixing a regime-finding methodology and backtesting it is excellent. Otherwise, if one's identification of regimes is more discretionary, one's trading results would have to provide the empirical test of efficacy.

Thanks for the good question--


matw said...

TJ and Brett,

I am with you both as far as finding correlations, but I am stumped by back testing. I imagine that one would use a fairly well reasoned indicator in the first place. How does one come up with an independent back test? It seems to me that the forward process has used up the available information.


(A scientist hooked by the "Trade like a scientist" series of posts. I am a beginner, and I am learning and enjoying the process.)

Brett Steenbarger, Ph.D. said...

Hi Mat,

I believe you'd have to develop a mechanical process for selecting predictor variables and lookback periods and then apply the predictions to market results for testing. Few traders who work from regimes are so structured, though there are applications of data mining (such as the Salford System tools) that essentially take such an approach.


matw said...

Hi Brett,

Thank you. I had conceptualized the "forward" effort as including tests against historic data (i.e. learn from the past. I'm reading your book), but I'm gathering that a "back-test" is the process of testing a model against historic data.

Thanks again.