Adrian Reid explains walk forward optimization: iterative in-sample optimization followed by out-of-sample testing. It works well for short-term, high-frequency systems with many uniform trades, but fails for trend-following systems that trade infrequently with long, irregular holding periods. Reid uses it selectively based on system type.
Now, following from this, what are my thoughts on warp forward optimizer optimization? Warp forward optimization, pioneered by Robert Pardo is held up as the gold standard in back testing. And look, I say gold standard because that's what people say, right? And it is, it's the most rigorous way to test. Walk forward optimization is where you have your system rules and you test them over, let's say a year of data and then you, you find the best parameter combination and then you walk it forward six months on unseen data and you see how it performs and then you optimize it again on the next set of data and you walk it forward another six months and see how it performed on the unseen data and then you optimize it again and then you walk it forward and then you optimize it again and then you walk it forward and you're basically stringing together these walk these out of sample walk forward periods to develop the whole backtest equity curve and all of the performance stats. So it's quite rigorous in that you're not, you're not kind of overfitting and choosing the very best values and then saying, oh, look how nice and smooth my equity curve is. It gives you the ugly equity curves, the out of sample equity curves for you to make your decisions on. So it's quite good in that respect. However, there's a problem. It only works well for certain types of trading systems. If you have a very short term trading system, let's say a mean reversion system, and the trades are pretty highly uniform and you have many trades at each period in the backtest, walk forward can work really well because you get enough data in that small optimization window to get a good optimization value to test in out of sample as the optimization works. But if you've got a trend following system and it trades only a few times a year, you get a much smaller number of trades and the trades are long and they're irregular, there's a small number of really big ones and there's a lot of really small losses. It's really hard for walk forward optimization to generate sensible parameter values. So it just doesn't work well for trend following systems. So I use it for very short term stuff, I don't use it at all for trend following stuff. Sat.
