ನಿಫ್ಟಿ 50, ನಿಫ್ಟಿ ಬ್ಯಾಂಕ್ ಮತ್ತು ಸೆನ್ಸೆಕ್ಸ್ ಲೋಡ್ ಆಗುತ್ತಿವೆ…
Robustness research

Threshold Sensitivity

Does a tuned RSI level hold up? Each fold picks its threshold from training sessions only, then is scored on sessions it never saw.

Strategy Lab
This page exists to discourage threshold tuning, not to hand you a winner. Searching a grid on price history always finds something; what matters is whether it survives out of sample.
Sessions after a trigger that the return is measured over.
Walk-forward slices; more folds means smaller test sets.
What this page is for
What the page actually does

It takes one simple rule — RSI(14) crossing below a threshold — and tries every threshold on a grid, measuring the return a fixed number of sessions later. That part is what any grid search does, and it always finds a best cell.

Then it does the part a grid search does not: it splits the history into folds, lets each fold pick its threshold using only its own training sessions, and scores that choice on later sessions it never saw. Sessions whose forward window overlaps the boundary are purged and a further stretch is embargoed, so training cannot peek at the answer through the horizon.

The decision this supports — and how to make it

The decision is not "which threshold should I use". It is "should I trust a tuned threshold at all". Four numbers answer it, and any one of them can end the discussion:

  • Degradation wide → the grid was fitting noise.
  • p-value above ~20% → random draws did this too.
  • Selection stability below 50% → the folds disagreed, so there is no stable level to use.
  • Few observations → nothing was really measured, whatever the percentages say.

Only when all four look sound is there weak evidence — and weak evidence is the ceiling this page can reach.

What a pass would still not mean

It is one rule, on one listing, over one history, with no costs, slippage or position sizing anywhere in it. And the search you run yourself counts: sweeping twenty symbols until one passes is the same overfitting one level up, which is why the honest use of this page is to rule things out rather than to collect the ones that survive.

What each number means
Threshold / Fires / Positive — each grid cell: the RSI level tested, how many times the rule actually triggered across the whole history, and the share of those events that ended positive. A cell that fired twice tells you nothing, however good its mean return looks.
Out of sample — the headline honest number — the return earned by thresholds each fold chose from training data alone, scored on sessions it had never seen. This is the only column that resembles what you could have done.
In hindsight — what the best cell per fold returned, selected after seeing the answer. Unknowable in advance and printed only as the ceiling that hindsight buys.
Degradation — in hindsight minus out of sample — what the honest choice gave up. A wide gap is the signature of a grid fitting noise: the best cell was best by luck, and luck does not repeat on new sessions.
Permutation p-value — the returns are reshuffled at random many times and the whole exercise re-run. The p-value is how often chance matched or beat the real result. 33% means one random draw in three did this well — the result is ordinary. Low values are necessary but never sufficient.
Selection stability — how often the folds converged on the same threshold. A rule whose best level changes every period does not have a best level, so a low value here voids the exercise regardless of the returns.
Obs (tr/te) — training and test events in that fold. Test events are what the out-of-sample number rests on; a fold showing 2 / 0 scored nothing at all and contributes no evidence.
Purged + embargoed — observations dropped at each fold boundary because their forward window straddled it, plus a further gap held out after it. Without both, a fold's training data contains the answer to its own test.
Enter an exact exchange listing to sweep its RSI oversold threshold.