Quantitative Analysis to improve probability and results

[Carneros8]

By looking at your winners and losers in your strategy, you might have looked at attributes at the time of your trade execution for further analysis. For example, if you have a trend following strategy, you might jot down somethings like "quality of a trend", "volatility", "momentum", etc... and etc... upon execution of your order. Then you might look at over thousands of trades your strategy does and see what patterns exist for winners and losers thinking there might be something of value there.

However, you do some statistical research and you find that the same attributes that are seen in the winners are also present in the losers. So if you cut down the value to tailor to the losers, say if we adjust momentum down to 6, then we would cut 50% of the losers, but by doing so, you also cut down the numbers of winners as well.

So what do you realize about this approach to analysis and therefore your strategy? That perhaps your strategy is a pure probabilistic edge rather than a model...

 

[MarkBrown]

i think you post is more about people who use optimization to build models they know very little about the operation of it. i can see that they are generally just lost in a boom bust cycle of mush splitting data streams into half hoping a few waves big enough to overcome the ocean of chop.

trading is all about finding your edge, which will not be found in a pot of optimization alone. you have to first at least have a target idea and move on from there.

 

[Carneros8]

hey Mark

I can understand that the approach above can be geared towards a quant that purely relies on big data to find edge. However, it can also be a tool/method used for trade analysis. If someone wrote an automated strategy based on sound general principles, how would one attempt to improve the trade? They would probably measure some sort of attribute and maybe find some patterns (explanations) to see if there are overall patterns to the winners and losers.

It's interesting to think about strategies and thus your edge in two camps. One is a fundamental strategy maybe arb or some sort of quantitative model (Black Scholes, etc..), and the other is purely probabilistic, more winners than losers.

 

[smallfil]

If your expectation is negative, you can rely on your stats and you will still be a losing trader!
And if your statistics tell you, you will win 90% of the time, you will take it run with it! Not knowing that the 10% of the time that trade doesn't work, your loss is say 10-15 times your gain? We are not even talking about the times the statistics is skewed and you lose say, 5 trades in a row!

 

[Freestyle888]

By looking at your winners and losers in your strategy, you might have looked at attributes at the time of your trade execution for further analysis. For example, if you have a trend following strategy, you might jot down somethings like "quality of a trend", "volatility", "momentum", etc... and etc... upon execution of your order. Then you might look at over thousands of trades your strategy does and see what patterns exist for winners and losers thinking there might be something of value there.

However, you do some statistical research and you find that the same attributes that are seen in the winners are also present in the losers. So if you cut down the value to tailor to the losers, say if we adjust momentum down to 6, then we would cut 50% of the losers, but by doing so, you also cut down the numbers of winners as well.

So what do you realize about this approach to analysis and therefore your strategy? That perhaps your strategy is a pure probabilistic edge rather than a model...

I think you need to be careful about calibrating a model backwards through historical data (aka optimization), versus extrapolating that into a "probabilistic" forecasting model or "edge". An optimization approach is best suited for data validation of "model ideas" or "adjustments", but prone to mis-specification in the sense that it tends to reward specifications that fit that sample, and so no guarantee of out-of-sample. So lots of diagnostics and out-of-sample model mis-specification tests is required to avoid "corner solutions" that only fit in-sample. If your filter-rule impacts on both winners and losers, then its likely not to be very powerful, whereas a rule which only filters the losers, without zeroing in on one historical winner (that is not likely to be replicated), is likely to be "good".

 

[nooby_mcnoob]

I'm pretty sure this is a dumb question, but is there a way to set your risk/stop-loss based on win/loss probability?

 

[tiddlywinks]

I'm pretty sure this is a dumb question, but is there a way to set your risk/stop-loss based on win/loss probability?

If you are playing the "probabilities", a better question to ask is what is the trade with the highest probability and why should I trade anything else?

Yup, it takes work. But you also learn what your stop, or at least your uncle limit is. And this is very important... each and every time!

 

[Freestyle888]

I'm pretty sure this is a dumb question, but is there a way to set your risk/stop-loss based on win/loss probability?

I think you need a setup that looks more than just win/loss probability. That's probably too simple a characterization of your PnL function. You may need to consider things like maximum value-at-risk if using leverage, conditional factors that may affect the average win/loss ("the odds at the race-track were on my side until it rained, and then the odds didn't mean jack"), and also whether the strategy fits your psychological profile - can you really be a robot and strictly follow the rules and stop loss and take profits, or will you hang on a little longer to see how it plays out ...

 

[nooby_mcnoob]

If you are playing the "probabilities", a better question to ask is what is the trade with the highest probability and why should I trade anything else?

Yup, it takes work. But you also learn what your stop, or at least your uncle limit is. And this is very important... each and every time!

How do you do this without overfitting?

 

[nooby_mcnoob]

I think you need a setup that looks more than just win/loss probability. That's probably too simple a characterization of your PnL function. You may need to consider things like maximum value-at-risk if using leverage, conditional factors that may affect the average win/loss ("the odds at the race-track were on my side until it rained, and then the odds didn't mean jack"), and also whether the strategy fits your psychological profile - can you really be a robot and strictly follow the rules and stop loss and take profits, or will you hang on a little longer to see how it plays out ...

Oh definitely, this question is less about getting the most I can than it is about learning whether the setups I create follow the expected probabilities. Calibrating my process first. Does that make sense?