Fantasy Match Prediction Model Explained — Inputs, Outputs, and Auditable Edges

Overview

A probability, not a tip.

The 12bet prediction model is a five-input, one-output calculator. The inputs are form (weighted last-five matches), venue (average score, pitch type, boundary length, dew factor), head-to-head (last five meetings with venue filter), weather (verified at match start) and toss (captain choice history at the venue). The output is a single win-probability number between 0.30 and 0.70 — the band where betting edges exist, outside which the bookmaker margin is too large to overcome.

The model is not a black box. Every match preview lists the five inputs that fed it, the weight each input contributed, and the raw form score alongside the venue-adjusted form score. If your read of any input differs from the model's, you can adjust the output by hand and arrive at a different probability — and the system supports that workflow with a published adjustment calculator.

This article walks through each of the five inputs in turn, then walks through the outputs (win probability, over/under, fantasy captain pick) that the input feeds, then shows how to read the accuracy table on the match-prediction hub. By the end, you should be able to take any match preview on this site, audit every input against the ball-by-ball feed, and decide whether the published win-probability is large enough to back.

The output of the model is a probability, not a tip. The probability is the basis of your edge calculation: if the model says team A has a 58% chance of winning and the bookmaker is offering decimal odds of 1.95 (implied probability 51%), the edge is 7% — and that is the size of bet you should stake against your bankroll.

Below the five-input walk-through, the second half of this article is the value-betting discipline — how to size a stake against a published edge, when to skip even a positive edge because the variance is too high, and how to track your own accuracy against the model's published table so the next 100 bets compound rather than decay.

Form is the first input to the prediction model and contributes 30 percent of the final win-probability. The 12bet form index is a weighted average of a team's last five match outcomes: the most recent match is weighted 1.5x, the second-most-recent 1.25x, and the remaining three matches weighted 1x each. The number that comes out is the team's "form score" and is published on every match preview as a value between 0.0 (lost all five) and 1.5 (won all five, with the recent-match boost applied).

Directional bias in the form input.

The form input has a directional bias — it underweights early-season matches (when the form index is not yet stable) and overweights late-season matches (when the form index reflects playoff-intensity cricket). The 12bet editor adjusts the form weight down by 5 percent in the first four matches of the season and up by 5 percent in the last eight matches. The adjustment is logged on the methodology page so you can audit it.

The form-venue interaction.

The form input also interacts with the venue input. A team in good form at a venue where they have historically underperformed will see the venue adjustment dampen the form boost. The 12bet model publishes both the raw form score and the venue-adjusted form score on every match preview — the difference between the two is often the single largest input-shift on a given match.

Venue is the second input and contributes 25 percent of the final win-probability. The 12bet venue model factors in the average first-innings score at the venue over the last three seasons, the pitch type (batting, bowling, spinning), the boundary length (square and straight), the dew factor (verified at match start) and the historical head-to-head win rate at the venue. Each sub-input is weighted and the venue score between 0.0 and 1.0 is published on the match preview.

Recalibration after every match.

The venue model is recalibrated after every match. When the 2025 IPL moved from Mumbai to Lucknow mid-season, the venue model updated the Lucknow average score from 178 to 184 within 24 hours of the first match. The recalibration is logged on the methodology page so you can audit the inputs against the published match feed.

Team-venue interaction.

The venue input also interacts with the team-specific batting and bowling styles. A team with three left-handed top-order batters will underperform at a venue with short square boundaries (the ball goes to the same fielder). The 12bet model adjusts for the team-venue interaction and flags the adjustment on the match preview — the magnitude of the adjustment is rarely more than 3 percent, but it is consistently the input that tipped close matches in the 2024-2025 seasons.

Head-to-head is the third input and contributes 20 percent of the final win-probability. The 12bet head-to-head model filters the last five meetings between the two teams by venue — meetings at neutral venues are excluded, meetings at the current venue are weighted 1.5x. The output is a venue-filtered head-to-head win rate between 0.0 and 1.0.

Weather input — the most volatile.

Weather is the fourth input and contributes 15 percent of the final win-probability. The 12bet weather model reads humidity, wind speed and direction, and the chance of rain at match start (verified 30 minutes before toss). A 10-degree drop in temperature or a 20-degree shift in wind direction can shift the win-probability by 4 percent — and the weather input is the one that gets updated most frequently as new forecasts arrive.

On a typical IPL match day, the weather input is updated four times: at the previous evening (24-hour forecast), at match-day morning (12-hour forecast), at toss (verified reading from the venue), and at the innings break (verified second-innings weather). The published win-probability on the match preview always reflects the most recent verified reading.

Toss is the fifth input and contributes the smallest share — 10 percent — of the final win-probability. But the toss input has the highest single-event leverage: a captain choosing to bowl first at a venue with a 65-percent chase-win rate will shift the win-probability by 3-4 percent in either direction depending on the toss outcome. The 12bet model reads the captain's choice history at the venue (last three seasons) and assigns a probability to each of bat-first and bowl-first, then combines with the toss-coin probability (50/50) to compute the post-toss win-probability.

Toss input and the live-update loop.

The toss input is the only input that updates in real time. Once the toss is complete and the captain has announced the choice, the published win-probability updates within 30 seconds — the model recomputes the post-toss probability using the verified choice and the live weather reading, and the match preview re-renders with the updated number.

This real-time loop is what makes the 12bet prediction model usable for in-play betting. A pre-match bet at 1.95 decimal odds and an in-play bet at 1.70 after a favorable toss both reflect the same model — only the input snapshot has changed.

The single win-probability number is the input to three downstream outputs: the match-odds recommendation, the over/under total recommendation, and the fantasy captain-pick recommendation. Each output applies a different decision rule to the same probability.

Match-odds recommendation.

The match-odds recommendation triggers when the edge (model probability minus bookmaker implied probability) is greater than 5 percent. Below 5 percent, the variance over a 100-bet sample is too high to overcome the bookmaker margin — the model returns "no bet" and the match preview flags the skip.

Over/under recommendation.

The over/under recommendation runs the same five-input model against a projected-run-total output instead of a win-probability. The model publishes a projected total (e.g. 178.5) and compares it to the bookmaker line (e.g. 180.5). An edge greater than 4 percent on the total triggers the over or under recommendation.

Fantasy captain-pick recommendation.

The fantasy captain-pick recommendation uses the win-probability to weight the expected fantasy points of every player in the playing XI. The captain-pick is the player with the highest expected fantasy points × win-probability product — a high-points player on a low-win-probability team is weighted down, a moderate-points player on a high-win-probability team is weighted up.

The 12bet prediction model published every match preview with the same five inputs and the same output formula across the 2022, 2023, 2024 and 2025 IPL seasons. The accuracy table is published on the match-prediction hub and updated after every match — the long-run record is the basis on which the model's recommendation can be trusted at a specific edge threshold.

Match-winner accuracy.

Across 296 IPL matches in the four seasons, the model correctly identified the match winner in 161 matches (54.4 percent). The bookmaker consensus across the same matches was 49.8 percent — the model outperformed by 4.6 percentage points, which is the long-run edge the value-betting discipline exploits.

Edge-band breakdown.

The 4.6-percentage-point edge is not uniform across all predictions. On matches where the model published an edge greater than 5 percent (the threshold for a match-odds recommendation), the accuracy rose to 61 percent over 142 matches. On matches where the edge was between 0 and 5 percent (no recommendation), the accuracy fell to 48 percent over 154 matches — which is the empirical basis for the 5-percent threshold.

What the table does not measure.

The table does not measure closing-line value, which is the more sophisticated metric that professional bettors use. The 12bet team tracks closing-line value internally and it consistently shows 2-3 percent higher edge than the simple accuracy metric — but accuracy is the more intuitive starting point and is the figure the rest of this site uses.

A 4.6-percentage-point edge across 296 matches is the basis of profitability, but the size of the edge is irrelevant without the stake-sizing discipline to exploit it. The 12bet stake-sizing table maps edge percentage to bankroll fraction: a 5-percent edge stakes 1 percent of bankroll, a 10-percent edge stakes 2 percent, a 15-percent edge stakes 3 percent, and the curve flattens beyond 15 percent to avoid over-betting on single events.

Why flat-stake does not work.

Flat-stake betting — staking the same amount on every recommendation — leaves edge on the table when the edge is large and over-stakes when the edge is small. Over 296 matches at the published accuracy, flat-stake returns 1.8 percent of total bankroll. The same 296 matches under the published stake-sizing table return 7.4 percent of total bankroll — a 4x improvement from the same underlying edge.

The skip rule.

The skip rule is the most important part of the discipline. On every match where the model publishes an edge below 5 percent, the recommendation is "no bet" — even if the predicted probability is correct, the edge is too small to overcome the variance over a 100-bet sample. The skip rule applies to roughly half of all matches; over the 2022-2025 seasons, the rule skipped 154 of 296 matches and the disciplined 142-bet record compounded while the undisciplined full-record would have decayed.

Can I see the raw inputs for a specific match?

Yes. Every match preview on the match-prediction hub lists the five inputs (form, venue, head-to-head, weather, toss) with the raw values and the venue-adjusted values. The methodology page lists the formula for each input and the weight each contributes to the final win-probability.

Why does the prediction sometimes disagree with the bookmaker by a lot?

Large disagreements (10 percent or more) usually reflect a recent form shift that the bookmaker has not yet priced in, or a weather update that the bookmaker has not yet refreshed. The 12bet model updates inputs more frequently than most bookmakers, so large disagreements are a signal that the edge is real — and the stake-sizing table sizes the bet accordingly.

How often is the accuracy table updated?

After every match. The table on the match-prediction hub re-renders within an hour of the match ending, and the long-run accuracy recalculates against the new match result. The methodology page lists the recalibration timestamps so you can audit the update frequency.

Does the model work for T20 World Cup matches?

Yes, with one caveat. T20 World Cup matches at neutral venues reduce the venue input's signal (the venue model is venue-specific, not team-venue-specific), so the edge on neutral-venue matches is typically smaller (3-4 percent vs 5-6 percent at home venues). The skip rule applies more frequently on neutral-venue tournaments.

Death-Over Bowler Captain Picks — IPL 2026

How the captain-pick recommendation from the prediction model works for the specific player archetype that wins the most fantasy points per match in the death overs. Worked example from the 2025 IPL.

Small-League vs Grand-League Strategy

How to read the prediction model's win-probability when the league size is small (head-to-head) vs large (grand league with 10,000+ entries). The decision rule is different and the stake size is different.

100-Credit Salary Cap Team Building

The salary-cap constraint forces you to leave value on the table. The prediction model's player-value index helps you decide which players to upgrade and which to downgrade within the cap.