How the engine is checked on every commit
A backtest tool is only worth trusting if its arithmetic is exactly right, and if running the same session twice gives the same answer. QKA's kernel — the deterministic reducer that turns a stream of events into simulation state — is guarded by three stress scripts that must pass before any kernel change ships:
| Script | What it drives | What must hold |
|---|---|---|
| stress_determinism.py | 100+ randomized sessions, 5 replays each | Byte-identical outcomes every replay |
| stress_time_engine.py | Randomized Time Engine sessions | Live final state == event-log replay |
| stress_trading_engine.py | Randomized trading sessions | Live state == replay; P&L reconciles exactly |
These run on randomized synthetic sessions, not a specific market's price history — they prove the engine's bookkeeping is exact: the same events always produce the same state, and the live simulation never silently drifts from what its own event log says happened. That property doesn't depend on which market's candles are loaded into it.
The rigor toolkit — what it's built to catch
A correct engine can still flatter a bad idea if it only reports one number. The Research Desk's job is the opposite: to say "this looked good, but it won't repeat" when that's true. Four tools exist for exactly that:
Overfitting gate
- Re-scores the best of many tested variants out-of-sample against a no-skill baseline, so an inflated in-sample winner gets caught before it's trusted.
Walk-forward split
- Re-runs a rule on the earlier and later halves of a window separately; a rule strong early and weak late is flagged as overfit rather than reported as one blended number.
Bootstrap resampling
- Resamples a rule's actual trades to build a confidence band, so a result resting on a handful of trades reads as fragile rather than solid.
Out-of-time forward test
- Freezes a rule's parameters before evaluating it on data the selection process never touched — the only fully honest check against learning-the-past-not-the-future.
These four tools were built and exercised end-to-end in July 2026, before QKA moved to a NEPSE-only product. That study is preserved in the project's internal records — it isn't reproduced here, because it evaluated market data this product no longer trades, and its figures would describe that market, not NEPSE.
What this page does and doesn't claim
What's actually verified
- The deterministic kernel's bookkeeping is exact: live state and event-log replay always agree, checked on every kernel-touching commit.
- The overfitting, walk-forward, bootstrap, and forward-test tools exist and have been exercised end-to-end at least once.
- The product trades NEPSE equities exclusively — real Nepal Stock Exchange market data, zero cryptocurrency, since 2026-09-09.
What isn't claimed
- That any specific NEPSE trading rule is profitable, or has been backtested and published. No such study exists yet.
- That an earlier study run before the NEPSE migration describes NEPSE performance — it doesn't, and its figures aren't reproduced on this page.
- That backtest history predicts the future, on any market. The forward-test tool exists precisely because it doesn't.