How to read a model track record
Four habits for reading any model's record without being fooled by it — using the denominator, the drawdown, and a single committed call.
You do not need to audit a whole history to know whether a model is honest. Read one record carefully and confirm one call end-to-end, and you learn the thing that matters most: whether the record can be checked at all. The four steps below go from the cheapest, fastest screen to the single decisive check.
1. Start with the denominator
Find the total decision count and confirm the losers are included. A win rate with no number of calls behind it — or with the losses quietly removed — fails before you read another line. A model's 70% across 690 decisions says more than a 95% banner across an unstated few. This is the measurability test in its quickest form.
2. Find the drawdown
The worst peak-to-trough fall is the number that tells you whether the return was survivable. A record that shows a return and hides the drawdown has shown you the reward and concealed the risk. If you could not have sat through the dip, the headline return was never really yours.
3. Check that the score fits the clock
A fast model and a slow one should not be judged against the same bar. Confirm that each model is measured on its own returns — a grade should mean “above-typical for this clock,” not one absolute target stretched across situations it never fit.
4. Confirm one call before its outcome
This is the decisive step, and the one most records cannot survive. Take a single past call and match its published entry, target, stop and grade against its public, anchored receipt. Because the receipt was written before the trade resolved, a match proves those fields were fixed in advance — which is the one thing a curve-fitted backtest can never show. Here is exactly what that looks like:
The call below is a made-up illustration for the walkthrough, not a specific real trade. The procedure is exactly what you would run on a genuine published call from any of the four models.
- Take the published call and its fields. Say it reads: long,
entry 73.40,target 76.10,stop 72.20,grade B,signal time 09:48 UTC. - Reconstruct the fingerprint. The publisher concatenates those exact fields in a fixed order and runs them through a one-way hash — a function that turns any input into a single fixed-length fingerprint. The same fields always produce the same fingerprint; one changed digit produces a completely different one.
- Open the receipt. The receipt published with the call points to the public-ledger block its fingerprint was anchored in. Confirm the fingerprint you reconstructed matches the one in the receipt.
- Check the clock. Look up when that block was written. If it sits before the trade resolved, the whole call — entry, target, stop and grade together — was provably fixed in advance. That is the entire proof.
Try to break it: imagine the stop was nudged from 72.20 to 72.80 after price turned. Step 2 would then produce a fingerprint that no longer matches the receipt from step 3, and the tamper is exposed. That is why a confirmed receipt is worth more than any screenshot — it fails loudly the instant a field is touched.
Net: steps 1–3 take a couple of minutes and screen out most of what you will be shown; step 4 is the one that cannot be faked. The tested models recommended here support every step; the reasoning is laid out on the what-a-model-is page, and the property underneath step four is repeatability.