Public validation · predictiq_stack_v7

Don’t trust our probabilities. Check them.

Every number here is scored against finished matches we forecast before kickoff: not a backtest, not a training-set score. The raw figures are on a public endpoint with no key, so you can reproduce every one of them in a few lines of code before you write a line of integration.

Headline results

Matches scored

3,445

finished, forecast pre-kickoff

Log loss

1.0228

lower is better

Brier score

0.2042

lower is better

Calibration error

1.9%

avg gap, forecast vs reality

Last evaluated 18 Nov 2026, 15:08 UTC · refreshed automatically as results come in

How the record is kept honest

Frozen before kickoff

The forecast is recorded ahead of the match and never edited afterwards.

Scored on results

Only settled matches count. Log loss and Brier score punish confident misses.

Against the market

Bookmaker odds with the margin removed are the benchmark, because it is the hardest one.

Open to inspection

The same numbers come from a public, keyless endpoint. Reproduce them yourself.

When we say 60%, it happens about 60% of the time

This is the check that matters for a probability API. If forecasts were honest, each dot would sit on the dashed line. Dots drawn hollow have fewer than 30 matches and are not counted as evidence either way.

0%0%25%25%50%50%75%75%100%100%perfectHome win, bin 0.0-0.1: predicted 5.6%, happened 0.0% (n=3, too few to judge)Home win, bin 0.1-0.2: predicted 18.6%, happened 20.0% (n=15, too few to judge)Home win, bin 0.2-0.3: predicted 24.7%, happened 22.2% (n=423)Home win, bin 0.3-0.4: predicted 36.3%, happened 37.4% (n=991)Home win, bin 0.4-0.5: predicted 45.6%, happened 43.8% (n=949)Home win, bin 0.5-0.6: predicted 54.1%, happened 60.9% (n=908)Home win, bin 0.6-0.7: predicted 63.5%, happened 70.8% (n=113)Home win, bin 0.7-0.8: predicted 73.9%, happened 75.8% (n=33)Home win, bin 0.8-0.9: predicted 82.5%, happened 80.0% (n=10, too few to judge)Draw, bin 0.0-0.1: predicted 9.4%, happened 7.1% (n=14, too few to judge)Draw, bin 0.1-0.2: predicted 15.1%, happened 13.2% (n=159)Draw, bin 0.2-0.3: predicted 25.4%, happened 25.1% (n=3251)Draw, bin 0.3-0.4: predicted 31.2%, happened 14.3% (n=21, too few to judge)Away win, bin 0.0-0.1: predicted 7.1%, happened 9.1% (n=11, too few to judge)Away win, bin 0.1-0.2: predicted 17.0%, happened 13.9% (n=137)Away win, bin 0.2-0.3: predicted 24.6%, happened 23.6% (n=1723)Away win, bin 0.3-0.4: predicted 36.1%, happened 33.7% (n=999)Away win, bin 0.4-0.5: predicted 45.7%, happened 42.5% (n=315)Away win, bin 0.5-0.6: predicted 52.4%, happened 58.9% (n=236)Away win, bin 0.6-0.7: predicted 64.1%, happened 71.4% (n=21, too few to judge)Away win, bin 0.7-0.8: predicted 77.5%, happened 100.0% (n=1, too few to judge)Away win, bin 0.8-0.9: predicted 82.3%, happened 0.0% (n=1, too few to judge)Away win, bin 0.9-1.0: predicted 92.0%, happened 100.0% (n=1, too few to judge)Probability we forecastShare that actually happened
Home win Draw Away win under 30 matchesbar = 95% range, dot size = matches

Expected calibration error

  • Home win3.2%
  • Draw0.5%
  • Away win2.1%

The average distance between the probability we gave and how often it happened. Lower is better.

Every finished match's PRE-KICKOFF prediction is compared against the actual result. Calibration bins show, for predictions the model gave (say) 60-70% confidence to, what fraction actually happened - close to the bin's own range means the model's confidence is trustworthy, not just its picks. market_comparison scores the bookmaker's own de-vigged implied probabilities the same way, on the same matches, as a direct benchmark. roi_summary is a disagreement simulation, not a track record of profit: flat $1 stakes on fixtures where the model's probability exceeded the market's implied probability by a threshold, settled at average bookmaker odds. Small samples swing widely, so do not read it as evidence of an edge.
Show the bin-by-bin table

Home win

Confidence 0.0-0.1 (n=3) too few to judge
Predicted
6%
Actual
0%
Confidence 0.1-0.2 (n=15) too few to judge
Predicted
19%
Actual
20%
Confidence 0.2-0.3 (n=423) close to actual
Predicted
25%
Actual
22%
Confidence 0.3-0.4 (n=991) close to actual
Predicted
36%
Actual
37%
Confidence 0.4-0.5 (n=949) close to actual
Predicted
46%
Actual
44%
Confidence 0.5-0.6 (n=908) close to actual
Predicted
54%
Actual
61%
Confidence 0.6-0.7 (n=113) close to actual
Predicted
64%
Actual
71%
Confidence 0.7-0.8 (n=33) close to actual
Predicted
74%
Actual
76%
Confidence 0.8-0.9 (n=10) too few to judge
Predicted
83%
Actual
80%

Draw

Confidence 0.0-0.1 (n=14) too few to judge
Predicted
9%
Actual
7%
Confidence 0.1-0.2 (n=159) close to actual
Predicted
15%
Actual
13%
Confidence 0.2-0.3 (n=3251) close to actual
Predicted
25%
Actual
25%
Confidence 0.3-0.4 (n=21) too few to judge
Predicted
31%
Actual
14%

Away win

Confidence 0.0-0.1 (n=11) too few to judge
Predicted
7%
Actual
9%
Confidence 0.1-0.2 (n=137) close to actual
Predicted
17%
Actual
14%
Confidence 0.2-0.3 (n=1723) close to actual
Predicted
25%
Actual
24%
Confidence 0.3-0.4 (n=999) close to actual
Predicted
36%
Actual
34%
Confidence 0.4-0.5 (n=315) close to actual
Predicted
46%
Actual
43%
Confidence 0.5-0.6 (n=236) close to actual
Predicted
52%
Actual
59%
Confidence 0.6-0.7 (n=21) too few to judge
Predicted
64%
Actual
71%
Confidence 0.7-0.8 (n=1) too few to judge
Predicted
78%
Actual
100%
Confidence 0.8-0.9 (n=1) too few to judge
Predicted
82%
Actual
0%
Confidence 0.9-1.0 (n=1) too few to judge
Predicted
92%
Actual
100%

Our forecasts vs the betting market

On matches with bookmaker odds, the market consensus still scores slightly better than our early forecasts. We publish that too.

Same 2,865 matches, same scoring. Lower log loss and Brier score are better; a positive edge means we scored better than the market. Small gaps are within normal noise and should not be read as an advantage either way. Where odds exist, the published forecast follows the bookmaker consensus (margin removed), because our own model has not beaten it. For a developer that is the useful part: you get market-grade probabilities, plus a model for the matches the market does not cover.

Our log loss

1.0228

Market log loss

0.9579

Log loss edge

-0.0649

behind market

Brier edge

-0.0146

behind market

Matchday record

The forecast frozen about an hour before kickoff, which is the one you would use on the day. It is recorded once and never edited. 583 settled matches, average lead 56 minutes.

Accuracy

54.2%

Log loss

0.9532

lower is better

Statistical model on its own, same matches: log loss 1.0198, accuracy 48.9%. Where bookmaker odds exist, the published forecast follows the market, because our model has not beaten it.

  • Bookmaker consensus (margin removed): 572 matches, log loss 0.9543, accuracy 53.8%
  • Statistical model (no odds available): 11 matches, log loss 0.8955, accuracy 72.7%

Accuracy over recent evaluation runs

10 Oct 202618 Nov 2026

Verify it yourself

No API key, no signup. This is the same JSON this page is built from, so anything above can be checked against it.

Shell

curl -s https://api.predictiqpro.com/status/accuracy

Python

import requests

r = requests.get("https://api.predictiqpro.com/status/accuracy", timeout=20).json()
print("matches scored :", r["n_matches_evaluated"])
print("log loss       :", r["log_loss"])
print("brier score    :", r["brier_score"])

mc = r.get("market_comparison")
if mc:
    print("market log loss:", mc["market_log_loss"])
    print("edge vs market :", mc["log_loss_edge_vs_market"])  # negative = behind

# Worst calibration gap among bins with enough matches
gaps = [
    (abs(b["predicted_avg"] - b["actual_rate"]), side, b["bin"], b["n"])
    for side, bins in r["calibration"].items()
    for b in bins if b["n"] >= 30
]
print("largest gap    :", max(gaps) if gaps else "n/a")
Also public: /status/accuracy/history for every evaluation run, and /status/matchday for the forecast frozen about an hour before kickoff.

What this does and does not show

It does show

  • Whether our probabilities are calibrated: what we call 60% happens about 60% of the time.
  • How we score against bookmaker consensus on identical matches.
  • A record that grows daily and is published whether it flatters us or not.

It does not claim

  • That the model beats the market. Our own testing says it does not, and we say so.
  • That accuracy predicts betting profit. This is football data, not betting advice.
  • That small gaps mean anything: every figure has a sample size and a margin of error.
How to read each number
  • Log loss and Brier score: penalise confident forecasts that turn out wrong. Lower is better, and both are scored on the same matches as the market.
  • Calibration: the gap between the probability we gave and how often the outcome happened.
  • Accuracy: 49.3% of finished matches went the way we rated most likely (95% range 48–51%). It ignores confidence, so the three above matter more.

Disagreement simulation

A back-test, not a promise and not evidence of an edge: one unit staked on every match where our forecast beat the market’s by at least 3 points, at the market’s average odds. A result near zero is what our own testing predicts. It ignores stake limits, odds movement and margins.

Simulated bets

1,219

Win rate

36.8%

Units staked

1219

Simulated return

-6.9%

before real-world costs

Build on probabilities you can audit

The same probabilities, expected goals and goals markets behind these numbers are available through a REST API and an MCP server for AI agents. 14-day trial, card required, nothing charged until it ends. Cancel before then and you pay nothing.