Before you trust the result.
An executable validation harness that recomputes predictive accuracy, tests it against baselines, walk-forward folds, leakage and calibration — then returns a verdict with an exit code.
$4,950 · one payment Via escrow first Source code only
Try the free Backtest Lie Detector — paste a prediction file, get a verdict.
Run it in your browser All free toolsThis is not a mass-market product. It is a specialised validation harness for teams who need to prove or disprove backtest claims before deploying capital.
Six things can make a validation result look excellent while saying nothing. Each one is met by a specific check inside the harness — if the file is honest, the check answers it; if it isn’t, the verdict comes back OVERFIT · exit 1.
A CSV with a claimed-accuracy column tells you what the vendor wrote. The predictions tell you what happened. Check 01 recomputes from raw predictions and ignores the claim.
Accuracy means nothing until it beats the replies “up every day”, “no change”, and a momentum rule computed from the same data. Check 02 runs the baselines and never fakes the ones it can’t compute.
The demo folds from 80.00% down to 55.83% while the headline reads 69.33%. Check 03 measures fold by fold, from earliest to latest, so decay can’t average out of the story.
Trying many configurations guarantees a lucky winner. Check 04 applies a Bonferroni correction on the trial count — the more you tried, the harder it is to pass.
Leakage quietly slips information the model couldn’t have had into the test window. Check 06 gates it — either the columns pass or the run fails, whatever the headline says.
Forty predictions at 70% is not four thousand at 70%, and vendors count on nobody asking. Check 05 computes Wilson 95% intervals on every accuracy figure the file can produce.
Feed it predictions. It does the rest — and the verdict comes with an exit code, so it can gate a pipeline instead of sitting in a report nobody reads.
Accuracy is recomputed from raw predictions. Any claimed-accuracy column in the file is ignored. If the file says 92% and the predictions say otherwise, the predictions win. This is the check the whole harness is built around.
The model is compared against baselines computed from the data supplied — not against a number someone wrote down.
Five baselines are defined. The verified demo run measured all 5 and beat 5/5 (best baseline: momentum-5d at 51.00%).
Without price history (e.g. --no-prices, or the market service unreachable) only 2 run: always-bullish and no-change — the other three are never faked.
Accuracy is measured fold by fold, earliest to latest, so decay shows up instead of hiding inside one average. This is period stability: how the model holds up across time periods. Purged walk-forward exists as a tested library function, but it is not a DIRECTION verdict gate. The distinction is stated here because buyers should not have to discover it later.
Trying many configurations and reporting the best one is corrected for — the more you tried, the harder it is to pass. Bonferroni correction is applied to the number of configurations tried. That count is supplied with the run, so it is only as honest as the input — the harness does not pretend to know what it wasn’t told.
Every accuracy figure carries a Wilson interval, so a small sample can’t dress up as certainty. A 70% on 40 predictions and a 70% on 4,000 predictions are not the same claim, and the output doesn’t let them look like one.
Checks for leakage — information in the test data that the model shouldn’t have had when it made the prediction. A model that “predicts” with tomorrow’s information fails here, however good its headline number looks.
When a candidates file is supplied, the harness estimates the probability that the chosen model only looks best because many were tried. For a single model with no candidates file, PBO is null — and never a gate. It answers one question: among everything you tried, how likely is this winner a fluke?
Checks whether the model’s stated confidence matches how often it is actually right. A model that says 90% and is right 60% of the time is miscalibrated — useful to know before its probabilities drive position size.
In the verified demo, the sample model starts at 80.00% and ends at 55.83%. The overall 69.33% hides the story. The folds tell it.
One average can flatter a model; a sequence cannot.
Two things transfer in one sale: the certifier itself, and the forecasting platform it was built against — as a reference implementation, not as a running business.
A CLI-shaped harness that returns a verdict and an exit code, so a bad backtest can fail a pipeline rather than a meeting.
A Next.js platform included so the certifier can be seen in context — pages, routes, models and all. It is not a business in a box.
models/.Every number on this page traces to one command over one fixture. Run it in the acquired repo — if the harness doesn’t return this verdict, don’t buy the tool.
npx tsx scripts/certify-model.ts examples/vendor-model-sample.csv --configurations 20
79efd015a8501d1fa7ba79931532f5683b4c0e5f9bb56263d2544a8877cb3bac
This section is the sale. If any of it is a problem, it should be a problem now — not after escrow.
Pre-launch. No customers, no accounts, no domain, no brand to inherit. There is nothing to “take over” — only code to own.
It measured ~49.7% directional accuracy at n = 1,160. You are buying the tool that proves that — not a predictive edge.
No uptime history, no live broker execution, no licensed market-data feed included. The platform has never served a real user.
Without price history a run is capped at 2 baselines, not 5 (the verified demo ran all 5 and beat 5/5). Walk-forward here is period stability, not purged out-of-sample training. PBO is null for single models. Purged walk-forward and deflated Sharpe are tested library functions, not verdict gates.
Development used AI coding tools extensively. Copyright in purely AI-generated expression is unsettled, and that is disclosed in writing to buyers.
The repository was public under an MIT licence from 1–22 October 2026. Anyone who copied it then holds a perpetual MIT licence to that snapshot. Exclusivity cannot be warranted, and the sale does not claim otherwise.
Nothing on this page is a footnote to be discovered later. The limits above are the offer, in full.
Read the termsThe sequence is fixed. Money is secured before code moves — for both sides.
The buyer funds escrow for the full $4,950. Nothing moves before this.
Full repository transfer at the agreed commit, as a squashed snapshot.
The buyer checks what was received against these terms.
Funds release only after confirmation. No repo access before escrow funds.
A direct conversation with the person who built the tool — no sales team, no broker, no marketplace cut.
Builder of BuildIT Index · Seller, directly
BuildIT is Dron Pancholi’s project, and this sale is direct: the person who built the harness is the person who answers your questions and hands over the code.
That is deliberate. A certification tool is only worth what its maker will stand behind in the open — including the limits listed above.
No agency, no broker, no marketplace cut layered on top. The terms on this page are the terms.
The questions a diligent buyer asks first — answered here, in the same words as the terms above, so nothing changes between this page and the paperwork.
The source code only: the certification harness plus the forecasting platform it was built against, as a reference implementation. No customers, no accounts, no API keys, no .env file and no licensed data feed are included.
The buyer funds escrow for the full $4,950 first. The repository is then transferred as a squashed snapshot, the buyer confirms what was received, and escrow releases. No repo access is given before escrow funds.
7–14 days of questions only. No SLA, no maintenance and no future updates. The escrow fee is paid by the buyer or deducted from the proceeds at close.
Yes. The repository was public under an MIT licence from 1–22 October 2026. Anyone who copied it then holds a perpetual MIT licence to that snapshot, so exclusivity cannot be warranted.
Setup is 3–5 engineer-days on Node 22 (the repo’s CI pins Node 22), running costs are $35–80 per month at hobby scale, and the Chronos-2 weights (~464 MB, Apache-2.0) download from Hugging Face at first run.
A free short summary is available to anyone on request. The full proof pack is shared privately, only after screening, under an evaluation-only licence with no commercial use or redistribution.
It measured about 49.7% directional accuracy at n = 1,160 — a coin flip. You are buying the tool that proves that, not a predictive edge.
A free short summary first if you want it. Serious buyers get the full proof pack after screening — nothing is sold on a first email.
Choose what you’d like, add a line about yourself if you wish, and send the enquiry.
Every enquiry is read by Dron directly. Ask for the demo, the summary, or the full terms — before anything is discussed about price, the facts on this page are the starting point.