Public accuracy ledger
Model performance & accuracy
Our forecast engine is a logistic-regression model trained on PSX end-of-day history. Retraining is scheduled each weekday evening, and a new model goes live only if it beats the old one on held-back data; the last run is shown below. Here is how it actually performs — measured against its base rate, not marketed.
Model logistic-v1 Horizon 10 trading sessions Ledger since Graded on PSX end-of-day closes
What this means
Of 4,564 directional calls graded so far, 58.6% went the way the model said, against 67.2% for always guessing the commoner direction. So far that is no better than the simple baseline.
58.6%
4,564 graded calls · baseline 67.2%
59.2%
1,398 calls at ≥60% confidence
57.1%
up +0.79 ppvs 56.3% baseline · 64,406 bars
▼ −0.16%
mean per graded call, same window
190,030
17 features · all-time history
1 day
ago · 27 Oct 2026 · live model from 24 Oct 2026
Accuracy over time
Weekly hit rate of published calls, and each training run's held-out and in-sample score. Dashed lines are the base rates each has to beat: always naming the commoner direction over the same windows.
Latest week: 65.5% of 1,671 calls, baseline 79.2%. Latest held-out: 57.1%. Latest in-sample backtest: 50.9% across 60 runs — an upper bound, not an expectation.
Calibration: stated confidence vs what happened
ConfidenceHow strongly the model leans toward its call. It is not the probability of being right — the graded record against the base rate is the figure for that. GlossaryIf the confidence figure means anything, calls made with more of it should come true more often. Graded published calls only.
| Stated confidence | Calls | Avg stated | Came true | Gap |
|---|---|---|---|---|
| Below 55% | 2,190 | 48.6% | 57.0% | ▲ +8.4 pp |
| 55–60% | 976 | 57.4% | 61.1% | ▲ +3.7 pp |
| 60–65% | 826 | 62.4% | 61.1% | ▼ −1.3 pp |
| 65–70% | 434 | 67.1% | 58.3% | ▼ −8.8 pp |
| 70–80% | 138 | 71.7% | 50.7% | ▼ −21.0 pp |
Gap is how far the share that came true sits above (▲) or below (▼) the confidence stated. Higher confidence has not meant higher accuracy so far. Read the confidence figure as a weak lean, not a probability.
Current model outputs
AI signalThe model's directional reading of the stock over the next ten sessions, on a five-step scale from strongly bearish to strongly bullish, built from price and indicator history. A description of the data, not investment advice. GlossaryHow the model's latest call is spread across the stocks it covers. A tally of outputs, not a recommendation.
How to read these numbers
The accuracy figure below is directional hit rate: how often the signal matched the direction the stock subsequently moved. It is measured out-of-sample — on data the model had never seen when it made the call — not on the history it was trained on. That distinction is the whole point. Any strategy can be tuned to look brilliant on the past, because the rules were chosen knowing what happened; that is overfitting, and it is the main reason promising models disappoint in live use.
Read every hit rate against its base rate, not against 50%. The base rate is how often you would have been right by always naming whichever direction turned out to be more common over the same windows — and because markets rarely split fifty-fifty, it is often well above 50%. A hit rate is an edge only when it beats that figure. Right now the model does not beat its base rate, so treat its readings as a description of the inputs, not as a forecast with an edge. It is not going to look like 90%, and a site claiming that is either measuring on training data or not telling you the truth. Even with an edge, a model reading is one input among several — it is not a machine that tells you what to buy.
A hit rate is also not a return. Being directionally right more often than the base rate says nothing about the size of the moves you caught versus the ones you missed, and it excludes brokerage, taxes and slippage — which, as our getting-started guide explains, are what quietly separate a good-looking strategy from a profitable one.
Per-stock records below are shown only where there are at least 20 resolved predictions. Below that, a high percentage is noise: a stock the model called correctly four times out of five has told you almost nothing about its fifth call.
Limitations
- Calls are graded on direction only, 10 trading sessions after issue: a bullish reading on a stock that rose 0.1% counts as right, one that fell 0.1% as wrong.
- Neutral readings are never graded — a call with no direction cannot be right or wrong — so the hit rate covers bullish and bearish readings only.
- Backtest and per-stock figures are measured on history the model was fitted to. Read them as upper bounds.
- The record covers one model over one stretch of the market. A policy shock or currency crisis can break what held before.
- These are model outputs, not investment advice. See our methodology and disclaimer.
Strongest per-stock track records
Directional accuracy on each stock's own backtested forecasts (≥20 resolved). Backtests flatter — treat these as an upper bound.