11:00WONFT 0-3
Pisa
NapoliAway Win
Moderate@ 1.3662%
Issue 142 · SUN 17 MAY
Predictions for 17 May.
13 predictions across 9 leagues. Confidence is calibrated — when we say 70%, we mean 70%. What do these numbers mean?
Today
13predictions
30-day acc.
76%calibrated
About these predictions
This page is the model's daily shortlist — the fixtures where a calibrated probability clears our confidence threshold for a given market, across every competition we cover. It isn't a tip sheet of hunches. Each prediction is a model probability, shown alongside the bookmaker's odds (model vs market) and a confidence rating, so you can see why a fixture made the slate rather than just being told to back it.
Markets we cover
We publish the markets people actually follow, each as its own calibrated probability:
- Match result— home win, draw or away win, summing to 100%.
- Double chance— home or draw (1X) and away or draw (X2), the safer two-of-three result cover.
- Both teams to score— whether each side scores at least once. The dedicated BTTS page lists these on their own.
- Over / under goals— the 1.5, 2.5 and 3.5 total-goals lines.
How the predictions are generated
Two Poisson regressors estimate each team's expected goals from recent venue-specific form, Elo ratings, rest days, head-to-head history and recency-weighted shot and expected-goals data. Those feed a joint scoreline distribution with a Dixon-Coles correction, from which the goal and BTTS markets are read off; the three-way result comes from a separate multinomial classifier that uses the same inputs plus the Poisson outputs. Every market then passes through an isotonic calibration step so the published numbers match real-world frequencies. The full pipeline is on the methodology page.
How to read confidence and model vs market
Each tip shows the model's probability, the current bookmaker odds, and a confidence rating reflecting how far that probability clears the market's threshold. “Model vs market” is a sanity check, not a value claim — the market is a strong baseline, and our models are built to useprices as context rather than to be second-guessed against them. Precise percentages are shown for signed-in users; everyone can see which fixtures qualified and how each market has performed over the recent past. These are probabilities, not certainties — even a 75% pick loses one time in four.
Frequently asked questions
- Are these football predictions free?
- Yes. The daily slate is free and updated every day. A free account unlocks the exact model probabilities and confidence ratings on each fixture; the public view shows which fixtures qualified and how each market has performed recently.
- What markets do you predict?
- Match result (home win, draw, away win), double chance (home or draw, away or draw), both teams to score, and over/under 1.5, 2.5 and 3.5 goals. Each is a separately calibrated probability rather than a single combined rating.
- What does “model vs market” mean?
- It compares the model's probability for a fixture with the probability implied by the bookmaker's odds. It's a sanity check and a way to see where the model and the market disagree — not a signal to bet purely because they differ. We tested an explicit model-beats-market value selection and retired it: on our data the market price out-predicted the model on exactly the fixtures where the model deviated most, so disagreement was mostly model error, not edge.
- How is the confidence rating calculated?
- Confidence reflects how far a fixture's calibrated probability clears the on-page threshold for its market. A pick that only just qualifies scores lower than one well above the line. It's a margin-of-safety indicator, not a separate prediction.
- How often is the slate updated?
- Daily. Predictions are published ahead of each day's fixtures, and once a match finishes its row settles to won or lost with a tick or cross. Use the date pills to move between recent days and tomorrow.
- Do you guarantee winning predictions?
- No. These are calibrated probabilities, not certainties — a fixture shown at 70% is still expected to lose about three times in ten, and that's the model working as intended. We publish a rolling track record so the calibration is checkable rather than asking you to take it on faith. Please gamble responsibly.