5 · Sensitivity

Roadmap

Is your edge a broad plateau or a knife-edge? Robust systems degrade gracefully.

On the roadmap — not shipped yet

This stage is designed but not yet built into Malgot. The page below teaches the methodology so you can practise it manually today and recognise the feature when it lands. Anything describing a concrete Malgot screen is marked accordingly.

Optimization hands you a best parameter set. Sensitivity analysis asks the more important question: what happens to performance when you're slightly wrong about those parameters? A trustworthy system sits on a broad plateau — nudging a knob barely moves the result. A fragile one sits on a spike — the "best" setting is a statistical accident and its neighbours are losers.

Plateau, not peak

Prefer the middle of a plateau over the tip of a peak

A slightly lower backtest score in a wide flat region beats a slightly higher score on a lonely spike. The plateau will still be there next year; the spike won't.

The reason is simple: your parameters are estimates, and markets drift. If moving entry_z from 1.25 to 1.4 collapses the return, then the real world — which never hands you exactly 1.25 — will collapse it for you.

How to test it

One-at-a-time sweeps

Hold everything fixed, vary one parameter across its range, and plot the objective. You want a smooth, flat-topped curve — not a single tall bar.

Neighbourhood check

Around the chosen point, perturb each parameter by ±10–20% and confirm performance stays acceptable. If several small perturbations combine to wreck it, that's a red flag.

Heatmaps for interactions

For two interacting knobs, a 2-D heatmap of the objective reveals whether the good region is a connected plateau or scattered islands (islands = luck).

Relationship to the other stages

Sensitivity is the natural companion to Optimization: the optimizer finds candidates, sensitivity tells you whether a candidate is a real region or a mirage. It differs from Robustness, which perturbs the world (costs, data, regimes) rather than your parameters.

In Malgot

Roadmap

Not yet a dedicated feature — but it falls out of the optimization work almost for free, because a grid/random sweep already produces the objective at many parameter points. The intended surface is a 1-D sweep plot and a 2-D heatmap over the optimization results.

Practise it today, manually: run a small series of Quick Sims varying a single parameter in even steps, and record the objective for each. Even five points reveal whether you're on a plateau or a cliff. The histogram view on each run helps you confirm the shape of returns is stable across the sweep, not just the average.

4 · Optimization

6 · Robustness