Usually even good performing quant trading strategies work for a while and then return start to shrink. I see two reasons for that which would probably give rise to different analysis:

  1. The Strategy got known by too many traders and has been arbitraged away.
  2. Market conditions have changed (will or will not revert).

3 Answers 3


I go out on a limb and say No.

You can of course observe how it does, but making a prediction about how and when it decays is difficult to impossible with any degree of precision. You'd need a meta-model of the market as a whole. And, well, if you had that, wouldn't you use that knowledge to make your model better?

That said, you can of course measure pnl and other return characteristics and extrapolate, but that isn't a proper predictive model in my book.

  • $\begingroup$ Dirk, I am curious about how you would go about defining a meta-model of the market as a whole. If there 'exists' a meta-model 'out there', what features would it have? $\endgroup$ Feb 4, 2011 at 17:00
  • 4
    $\begingroup$ That's my point, this was a rhetorical device. You can't, hence you cannot model how your trading model would do within the realm of your meta-model. $\endgroup$ Feb 4, 2011 at 17:04
  • 1
    $\begingroup$ @Vishal The existence of a meta-model sounds like a contradiction, like Russell's paradox. Actually, I'd be really interested if there were a proof of this, though I don't know of one. $\endgroup$ Feb 4, 2011 at 18:31
  • $\begingroup$ Thanks. I got the 'impossibility' or proof-by-contradiction sense of it, but I was hoping to extract technical wisdom from Dirk :-) $\endgroup$ Feb 6, 2011 at 15:34

there could be a simple answer;

observe the drawdown in backtesting results; double it to get a conservative estimate.

if your model has exceed the theoritical drawdown; well; your strategy is breaking up with you.


The rule of thumb in one large derivatives group in the mid-90's was "about 3 months" before arbitrage starts causing serious damage. One place to start might be to look for other (anecdotal or survey-based) timeframes like that, see if you can get any sort of curve, trend, or surface out of that raw data.

But I suspect you might find that the valid half-life of any given model has more to do with overall market conditions than with the model itself. I also suspect that you might find a strong correlation between model half-lives and the shape of the yield curve.

Restating as a wild guess: If a model accurately describes some part of the market today, then it will likely do so tomorrow as well, with a probability of P. Assume that there is some relationship between P and the yield curve. For a first approximation, if your model is working with 2-year instruments, then use some factor multiplied by the 2-year yield curve slope to get P, and so on.

Using the yield curve to get P might work better with macroeconomic models, and not so well with micro. There are some obvious cases where this won't work -- if the model depends on some HFT or microstructure feature such as an exchange or counterparty's server load factor, for instance. A half-life rule for micro might be able to use something like VIX as an input though.

Again, this is all wild guesses, I've done little research to see if anyone else has written any papers about this. But a google search for "economic market model half-life yield curve" does look promising.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.