macrocephalopod

With some additional robustness checks and careful implementation of a simulator for fill prices, commissions etc you are well on your way to launching your first quant strategy.

Only problem is that the price series is one I generated using 100% random noise, it is completely unpredictable by any signal. Congratulations, you now know how to overfit a backtest. Welcome to quant research.

This is awesome, I like the idea of plotting sharpe as a function of parameter setting to find the sweet spot... good to bet on multiple settings tho for robustness. Since this series randomm,Any xtra steps u do to scrutinize whether it works bc of luck? Or is that always a risk

I mean this is why you have out of sample testing.

but if you have a thousand such time series, some of these will have good out of sample results. that’s really the whole game I think, balancing signal decay/transience vs random noise issues. it really helps to have a fundamental/structural view to weed out unlikely signals

this is why bonferroni exists and why parameter grid search is satanic

If you dont mind could you give a quick example for time series? Google isnt yielding great results on this. Suppose signal of interest is something Naive like RSI > 80, and hypothesis is that it causes abnormally low future 10 increment returns, any idea how to bonferri it?

I mean, simply, bonferroni would be used if you did a parameter search, e.g. if you checked RSI > 80, > 90, 10, 20, 30 increments. This is massively prone to overfitting. So you create a testable hypothesis - e.g. difference of mean returns for example (ideally bootstrapped)