Can you trust
that backtest?
A high-performing backtest is usually just fitted to past data. TestMyLogic is a desktop app that rigorously tests your crypto strategy's backtest for overfitting — check it before you trust those returns.
The app runs the statistics. Your AI agent explains them.
- No orders
- No exchange keys
- No fund movement
- Your data stays local
Hundreds of combos backtested, checked three ways
Three statistics each ask, in their own way, whether the backtest is real.
Puts the dozens to hundreds of backtests a sweep produced on a single map, and shows where the combination you pick ranks on eight measures.
How it works
Download market history
Pull public OHLCV and funding history into a local workspace.
Run a parameter sweep
Pick a strategy and a parameter grid; it backtests all of them.
Analyze the results
The backtest screen shows how robust your pick is; PBO and walk-forward each deliver a verdict.
Ask the AI agent you already use
We do not ship an agent. Connect the desktop app you already have — Claude, ChatGPT or Antigravity — and you can drive all of it by asking: download data, run the sweep, read the verdicts, query results with SQL, move the app's screens. Ask it to explain any term or number in plain words, or to compare the runs and point out what held up.
You: Download BTC, run a backtest sweep with the EMA_Crossover strategy, and analyze the results.
Agent: There was no stored history, so I pulled the whole BTC daily series — 3,317 rows, no gaps — and ran all 58 combinations over the ranges EMA_Crossover declares, on the app’s period plan: the last 3.0 years (2023-09 to 2026-08) are held back as a test segment, and every combination is picked on the 5.4 years before it. The development segment’s best returner is f10_s28 at +1,285%, but it spikes well above its neighbours and has only 28 trades there, under the 30 the validations require, so they leave it out. The combination the validation picks is f12_s16 — the most frequent winner across every possible split — and the optimum sits inside the grid, so there is no need to widen the ranges.
On the 3.0-year test segment those combinations had never seen, the walk-forward re-picked the parameters every window (training 270 days, validating 30, stepping 30, 36 windows). In the market only 55% of the time, it drew down -33.9% against -53.0% for simply holding BTC, and its annualised Sharpe of 0.98 beat holding’s 0.88. Its return, +123%, came in a little under holding’s +143%. Held fixed over the same days, the backtest winner made +121% and the most-picked combination +142% — and the most-picked one beat 68% of the other eligible combinations. Only 28% of them beat holding over those years, which is the market rising rather than anything about the selection.
The walk-forward verdict is “Not distinguished”: +123% sits inside the +86%–+173% range that picking a combination at random every window produces, and its average rank of #15 of 29 is inside the chance range of #13–#17. Cross-validation on the development segment agrees: PBO 71.4%, “Not distinguished”. That is not a bad grade — it says this data cannot show that picking the parameters beats picking at random, and more folds narrow the range. So: this strategy’s strength is its drawdown and its risk-adjusted return, the combination to carry forward is f12_s16, and the size to expect is the +142% the unseen years actually produced, not the backtest’s +1,285%.
Validate your backtest yourself
Pick your data and your strategy, and the statistics answer.
Download the free betaWindows 10 / 11 · 64-bit. The installer is code-signed — the download page shows the verified publisher.