How the SupaPicks model
selects every tip.
This page is a full technical explanation of the SupaPicks prediction system — how probability is calculated, how model edge is measured against bookmaker implied probability, how the optimal market is selected for each fixture, and how tips are assigned to products. We publish this because we believe bettors who understand their tools make better decisions. Nothing here is hidden.
The five-layer model at a glance.
The SupaPicks prediction model is not a single algorithm applied uniformly to every fixture. It is a five-stage evaluation process where each layer independently calculates a probability estimate for every possible outcome across every active market. The five estimates are then combined using a weighted ensemble that adjusts for recency, sample size and layer reliability by market type.
The five layers are: Expected Goals (xG) Data, Head-to-Head Base Rates, Contextual Motivation Analysis, Schedule and Fatigue Adjustment, and Market Signal. Each layer contributes a different type of evidence. A selection achieves high confidence only when all five layers converge on the same directional conclusion.
Expected Goals (xG) Data
xG measures the quality of chances created and conceded, corrected for shot location, type and context. It is a more stable predictor of future scoring than actual goals because it removes finish quality variance — a team that consistently generates 2.2 xG per home game will eventually score close to that rate even if their recent goal count is suppressed by goalkeeper performance or post-hits.
We calculate fixture-specific xG rates using a two-speed weighting: 60% weight on the last 8 games in the relevant venue (home or away) and 40% weight on the full season average. The home/away split is mandatory — a team's home xG rate and away xG rate are treated as distinct data series, never averaged together.
Opponent quality adjustment: each team's xG rate is adjusted by a strength-of-schedule factor calculated from the xGA (expected goals against) of their last 10 opponents. A team generating 2.0 xG per game against bottom-half defences is adjusted down when facing a top-three defensive side; a team generating 1.6 xG against top-half defences is adjusted up when facing a weaker defensive opponent.
The xG estimate feeds directly into a Poisson distribution model which generates the probability of every possible scoreline (0–0 through 5+–5+) for the fixture. Summing the probabilities of all scorelines where a given outcome occurs (e.g. all scorelines where Home team scores > Away team) produces the outcome probability from the xG layer.
Head-to-Head Base Rates
H2H base rates are the most fixture-specific evidence in the model — they capture structural patterns that persist across managerial changes, squad rebuilds and form cycles. Some fixtures produce draws at twice the league average; some produce Over 2.5 in 9 of 10 meetings. These patterns are not random — they reflect tactical match-up dynamics, rivalry intensity, and venue-specific factors that xG models cannot fully capture.
Critically, we track H2H data at the market level, not just the result level. For each fixture, we calculate: the BTTS Yes frequency, Over 2.5 frequency, over/under frequencies at all thresholds, HT/FT pattern distribution across all nine combinations, scoreline frequency distribution, and clean sheet frequency for each team. This market-level H2H data directly informs the relevant tip market rather than being routed through a match result estimate.
Sample size adjustment: fixtures with fewer than 6 H2H meetings in the relevant venue receive a reduced weight (down to 15% from 25%) and the remainder is redistributed to the xG layer. Fixtures with 10+ meetings receive the full 25% weight without penalty.
Contextual Motivation Analysis
Motivation analysis captures contextual factors that elevate or suppress performance beyond what xG and H2H data predict. It does not generate an independent probability estimate — instead it applies a multiplier to the combined xG/H2H estimate, shifting the output up or down based on the strength of the contextual case.
Positive motivation factors (increase attacking output): Must-win league position stakes; derby and rivalry context (adds +0.15–0.25 xG to the motivated side); elimination risk in cup competition; a team's star player returning from injury or suspension; manager under pressure with clear attack-first tactical directive; historical rivalry intensity index above league average.
Negative motivation factors (suppress attacking output): Dead-rubber end-of-season fixture with nothing to play for; known rotation ahead of a major cup final within 72 hours; a team that has already secured its primary objective (title, promotion, survival) with no secondary target; manager confirmed to leave at season end; tactical suppression directive (e.g. playing for aggregate score in European tie).
| Factor | Direction | xG Multiplier Range | Markets Most Affected |
|---|---|---|---|
| Must-win league position | ↑ Attack | +0.10 to +0.25 | Over 2.5, BTTS Yes, 1X2 |
| Derby / rivalry context | ↑ Both sides | +0.15 to +0.30 | Over 2.5, BTTS Yes, Offsides, Cards |
| Cup rotation ahead | ↓ Attack | −0.15 to −0.35 | All markets — both sides |
| Dead-rubber fixture | ↓ Both sides | −0.10 to −0.25 | Over/Under, Offsides, Corners |
| Playing for aggregate draw | ↓ Attack heavily | −0.25 to −0.45 | Under 2.5, BTTS No, HT/FT |
| Managerial debut | ↑ Effort | +0.05 to +0.15 | Cards, Fouls, 1X2 |
Schedule & Fatigue Adjustment
Schedule analysis quantifies the physical advantage or disadvantage created by the rest differential between the two teams. When one team has had significantly more time to recover since their last fixture, their physical output in the 60–90 minute window is demonstrably higher than average — and their opponent's is demonstrably lower.
Rest differential thresholds and adjustments: 0–1 day difference: no adjustment applied. 2 days difference: ±0.08 xG adjustment (rested team up, fatigued team down). 3 days difference: ±0.14 xG adjustment. 4+ days difference: ±0.20 xG adjustment, capped at this level regardless of larger rest gaps. These adjustments apply specifically to second-half performance metrics (applied to second-half xG estimates in the Poisson model, not total-game xG).
Additional schedule factors: travel distance for away teams (flights over 4 hours apply a −0.05 xG penalty); altitude differential above 1,500m; confirmed injury returns that increase available squad depth. All three are secondary inputs that modify the rest differential adjustment within a bounded range rather than operating independently.
Market Signal
Market signal tracks how bookmaker odds have moved from opening line to current price. Significant moves toward a specific outcome in the hours before kick-off indicate that professional or sharp money has been placed on that side — a meaningful signal that the market's consensus view of true probability has shifted.
Signal thresholds: A move of less than 5% in implied probability from opening: no signal. 5–10% move toward our model's selected outcome: weak confirmation (+5% confidence boost). 10–20% move: strong confirmation (+10% boost). 20%+ move: very strong confirmation (+15% boost, maximum). Moves away from our model's selected outcome are treated as a warning flag — we require re-examination of the other four layers before publishing.
Critically, market signal is a validation layer only — it cannot generate a new tip direction. A strong market move toward an outcome our other four layers do not support never results in a published tip. The signal layer can only increase confidence on a direction already supported by xG, H2H, motivation and schedule — it cannot override those layers in isolation.
How the five layers
combine into one number.
Each of the five layers produces a probability estimate for each market outcome evaluated. These are combined into a single weighted ensemble probability using the following base weights:
| Layer | Base Weight | Adjusted Weight (small H2H sample) | Role |
|---|---|---|---|
| 1 · xG Data | 35% | 50% | Primary scoring rate baseline |
| 2 · H2H Base Rates | 25% | 10–15% | Fixture-specific historical frequency |
| 3 · Motivation | 15% | 15% | Contextual intensity multiplier |
| 4 · Schedule / Fatigue | 10% | 10% | Rest differential adjustment |
| 5 · Market Signal | 15% | 15% | Validation and confidence modifier |
Measuring edge against
bookmaker implied probability.
Model edge is the difference between the model's true probability estimate (P_true) and the bookmaker's implied probability (P_implied), derived by converting the decimal odds to a probability and removing the bookmaker margin (overround) proportionally across all outcomes.
We publish a tip only when Edge ≥ 0.04 (4 percentage points). This threshold ensures that the tip has genuine expected value at the current bookmaker price and that the model's estimate is meaningfully above what the market implies — not just within the noise of normal modelling variance.
Choosing the highest-value market
for each fixture direction.
The model evaluates every market simultaneously for every fixture — it does not start with a match result direction and then find supporting markets. For each fixture, we calculate the edge across 1X2, DNB, DC, AH (−0.5, −0.75, −1, −1.25, −1.5), WEH, BTTS, Over/Under at each threshold, HT/FT, and player props where data is available.
The market with the largest positive edge is published as the tip for that fixture and direction. This is why our BTTS and WEH pages often carry tips on the same fixtures as our 1X2 page — at different market structures, different mispricing gaps appear.
The draw probability decision tree
When the model identifies a home or away direction, this rule governs which result-adjacent market is selected:
The four confidence
rating tiers.
Every published tip is assigned a confidence rating based on the number of model layers that converge on the selected outcome, the size of the model edge, and the H2H base rate validation. The four tiers map to the percentage chips displayed on every tip card:
How tips are assigned
to products.
Once a tip is generated and validated, it is routed to the appropriate product based on its confidence rating and model edge. The routing table below defines the threshold each product requires:
| Product | Min Confidence | Min Edge | Min H2H Rate | Tips / Day |
|---|---|---|---|---|
| Sure Wins | 85%+ prob. | 6pt+ | 8/10 | 5 |
| Sure 2 | V.High (92%) | 8pt+ | 7/10 | 2 |
| Sure 3 | High (80%) | 6pt+ | 7/10 | 3 |
| Genius Gold | V.High (92%) | 5pt+ | 7/10 | 8 |
| Daily VIP | High (80%) | 6pt+ | 6/10 | 4–6 |
| Standard market pages | Med (65%) | 4pt+ | 5/10 | 18 per market |
| Take the Risk | Any | 6pt+ | Any | 3–5 |
Take the Risk is the exception to the confidence threshold rule. It specifically selects high-odds markets (4.00+) where the model edge exceeds 6pt — regardless of the overall confidence rating. A correct score tip with a 22% true probability and 10.5% bookmaker implied is a +11.5pt edge tip at "Medium" confidence — too low for any other product, but exactly right for Take the Risk's value-hunting mandate.
The human layer
after the model.
Every tip produced by the model passes through an editorial review before publication. This is not a rubber stamp — analysts can and do override model outputs, though they must document their reasoning when they do.
The editorial review checks for factors the model cannot capture in automated data: late team news published after the model's last data ingestion (injury to a key player, unexpected lineup confirmation, manager illness); weather and pitch conditions that affect a specific match but are not reflected in historical xG data; referee appointments whose card-issuance and foul-tolerance profiles materially affect cards, fouls and offsides tips; and transfer window context where recent arrivals or departures have materially changed a team's tactical profile in ways the last-8-games xG data does not yet fully reflect.
Analyst overrides of model selections are flagged with an editorial note on the tip card and documented in the internal model performance log. Override decisions are reviewed monthly against actual outcomes to calibrate whether editorial intervention consistently improves or detracts from model accuracy. Based on the last 12 months of data: editorial overrides have improved tip accuracy by 2.1 percentage points net of additions and subtractions.
How we measure
model accuracy over time.
All tips are recorded in a central database at the time of publication with the following fields: fixture, market, tip direction, model confidence rating, model true probability estimate, bookmaker decimal odds at time of publication, model edge, product assignment, analyst who reviewed, and timestamp.
Settlement is recorded automatically at full-time for standard markets and manually within 24 hours for specialist markets (HT/FT, player props, correct score). Win rate, ROI and average odds are calculated on rolling 30-day, 90-day and 12-month windows and published on the Statistics page.
Calibration check: We regularly verify that our model's stated confidence levels match empirical outcomes. A correctly calibrated model's 92% confidence tips should win approximately 92% of the time over large samples. If V.High confidence tips are winning at 79% over 200+ selections, the confidence scale is over-stated — and we adjust downward. If they are winning at 95%, the scale may be under-stated. This calibration process runs quarterly and its outputs are factored into the next model version release.