[Hyperparameter Optimisation]

Search strategy input.* values over N trials (TPE / random / grid) with holdout or walk-forward. Strategies only.

Hyperparameter Optimisation

AXIS can search Pine Script™ input.* values on a strategy() script and apply the best assignment for this symbol and bar window.

The search engine is pyne (POST /optimize / pynescript.optimize). AXIS is the plugin UI. Pyodide and engines without /optimize fall back to an isolated local loop (random only).

This is not a live trading optimiser and not a Pine builtin. Best-on-sample is not a guarantee out of sample.

Warning

Holdout is the default for a reason. In-sample-only search overfits. Winners on one symbol or timeframe often fail on another.

Open it

Load a strategy and bars

Open a script that declares strategy(). Load history. A first Run is optional but helps engine-exported inputs appear.

Open Optimise

Results drawer → Optimise, or command palette Optimise strategy. Indicators and libraries show a one-line disabled reason.

Set the space

Tick the input.int / float / bool / enum fields to search. Numeric fields need min and max (minval/maxval from the script, or typed in the table). Source, color, timeframe, symbol, session, and text stay at their current Script Settings values.

Start

Set trials, sampler, objective, and validation. Start runs the study. Cancel aborts the in-flight engine call. Trials do not paint the chart.

Apply

Apply best merges the winner into Script Inputs (other inputs stay). Apply + re-run also runs the winner so the Strategy tab matches. Export the trial table as CSV.

Study controls

Control	Default	Notes
Trials	30	Hard cap 200
Sampler	auto	Random if N<20, else TPE (TPE needs pyne `/optimize`)
Objective	composite	Also net PnL, profit factor, Calmar-like
Validation	holdout 30%	Walk-forward or in-sample (warned)
Min trades	5	Fewer closed trades → trial rejected

Note

Grid is offered when the cartesian product fits the trial budget. On Pyodide (no /optimize) the UI falls back to random and says so.

Validation

Mode	What happens
Holdout (default)	Sampler sees the first ~70% of bars. Every trial is also scored on the last ~30%. Ranking prefers out-of-sample.
Walk-forward	Rolling train/test windows. Engine runs ≈ `N × folds` (capped at 400).
In-sample	One run on all bars. Warned in the UI.

Train and test are bar slices, not trade filters on a full-sample run.

Engines

Engine	Path
`server` (Pro API)	`POST {endpoint}/optimize` — parse-cached interpret loop, TPE / random / grid
Pyodide / missing `/optimize`	Isolated `runScript` — random only

Same-origin VPS / Docker must proxy /optimize to Flask (same as /run). Otherwise AXIS falls back to the slow client loop.

Overfitting

Prefer holdout. Treat in-sample net profit as a search score, not a live forecast.
Composite + min-trades rejects lottery-ticket parameter sets with 1–2 fills.
Apply + re-run on a later date range or another symbol before trusting a winner.