A Reproducible Walk-Forward Backtester

Most open-source backtesters give you an execution loop and leave the research discipline to you: how you split in-sample from out-of-sample, whether your indicators secretly peek at the future, and how an edge holds up once fees, slippage and funding are real. This framework moves that discipline into the engine. Walk-forward optimisation, regime segmentation, realism controls and a five-scenario robustness suite are built in, and a strict no-look-ahead invariant is enforced at the trade-ledger level so an author cannot accidentally trade on a bar that has not closed.

The same algorithmic specification is implemented twice. The Python reference is readable and easy to extend; the Rust port is built for speed. A parity harness compares their metric ledgers point by point, which is what lets the project claim its numbers are real rather than implementation lore. The figures below are generated from the framework's own runs.

A single backtest here is not a number, it is a dossier. One run of the bundled EMA-crossover strategy on SOL/USDT 1h emits an optimised in-sample and out-of-sample report, a Monte Carlo distribution over resampled trade sequences, a five-scenario robustness sweep, and a rolling walk-forward across eighteen windows. Everything in this section is the actual output of that run, drawn by the framework itself and shown unedited.

It is worth saying plainly: this strategy loses. That is the point. A bundled example that printed money would be the suspicious thing. What the framework is built to do is make a weak strategy's weakness impossible to hide, and the figures below are it doing exactly that.

The metrics it prints

Here is what the run reports. The optimised out-of-sample Sharpe is negative, and a Monte Carlo over resampled trade orderings ends in a loss on 99.8% of paths.

The robustness sweep re-runs the optimised baseline under four perturbations. Slippage shock is the worst case, roughly doubling the out-of-sample loss; a one-bar entry delay is nearly harmless. The spread between them is a cheap read on how fragile the result is to assumptions you cannot control.

And the rolling walk-forward, window by window. Some windows are positive (W01 forward-tests to plus $899), but a strategy that only works in some windows, on re-fit look-backs, is noise that the aggregate and the deflation gate are built to expose, not an edge.

Was it luck?

The Monte Carlo view resamples the trade sequence thousands of times and re-scores each metric, so you can see where the real result sits in the distribution of what randomness alone produces. When the actual value lands in the middle of a distribution centred on nothing, the result is indistinguishable from chance.