Methodology
This page explains how Proknoz generates its predictions. Transparency is a core principle — you should understand what the model does, what data it uses, and where it might be wrong.
Football — Dixon-Coles Model
Football predictions use a variant of the Dixon-Coles model, which extends the standard Poisson regression by correcting for the observed correlation between low-scoring outcomes (0-0, 1-0, 0-1, 1-1).
The model estimates attack and defence strength parameters for each team, along with a home advantage factor and a separate scoring level for each league — the Bundesliga and Liga Portugal do not produce the same number of goals, and a single average would understate one and overstate the other. European cup matches use the average of the league levels, because they bring together clubs from different leagues. National-team football has its own level. Every competition is fitted together on all matches since 2021, with an exponential decay that weights recent matches more heavily than older ones.
The model does not learn from goals alone. Goals are a noisy record of how a match went: a side can create three good chances, miss them all and lose 1–0. Where the expected goals (xG) of each individual match are available, the model is trained on a blend of 30% actual goals and 70% xG, so a team's rating reflects the chances it creates and concedes rather than how lucky its finishing was. In a backtest on the first months of the 2026/27 season this improved the over/under forecasts without changing match result accuracy. Matches without per-match xG (national teams, and older seasons outside the top five leagues) are trained on goals only. The probabilities themselves are still scored against real results. A regularisation term keeps these parameters anchored toward the population average, so a team only earns an extreme rating when the evidence genuinely supports it (see Calibration and Accuracy below).
At prediction time, four multiplicative adjustments are applied to the base expected goals (lambda) before computing match outcome probabilities:
- Recent form. A team's win rate over its last five matches is used to scale expected goals up or down by up to ±10%. A team on a winning streak attacks with slightly more confidence; one on a losing run is treated as weaker than its long-run parameters suggest.
- Rest days. Teams with fewer than four days since their last match have their expected goals reduced by 5% for fatigue. There is no penalty for a long gap: after an international break almost every team qualifies, so it would only lower the scoring level of a whole matchday.
- Injuries and suspensions. Absent players from Transfermarkt are mapped to canonical positions (Attacker, Midfielder, Defender, Goalkeeper). Each absence reduces the team's attacking or defensive expected goals according to the player's position and whether they are considered a key player. Suspended players are treated exactly like injured ones. An injury stops counting once the player's expected return date has passed, and absence data that has not been refreshed in the last 72 hours is ignored rather than trusted. The total adjustment is capped at ±25%.
- Expected goals quality. Season-long xG averages from Understat (top five European leagues) and SofaScore (Liga Portugal and the Championship) are compared to the league mean. A team that consistently creates more chances than its goals suggest receives a small upward adjustment; one that underperforms its shot quality is adjusted downward. This signal is deliberately light (±10% cap) to avoid competing with the attack/defence parameters already learned from historical goals.
Confirmed lineups. About an hour before kick-off, once both starting elevens are published, the match is predicted again. Each team's regular starters are identified from its previous confirmed lineups, and any regular who is not in the eleven counts as an absence in the same way as an injury — which catches rotation and late fitness decisions that injury lists miss. The pre-lineup forecast is kept, so the effect of the lineup can be measured.
Blend with the betting market. Where bookmaker odds for the match are available (the market-average pre-match prices from Football-Data.co.uk, usually published a few days before each round), the bookmaker's margin is removed and the expected goals that reproduce the market's match result and over/under 2.5 prices are calculated. The published forecast combines the model's expected goals with the market's, with 60% of the weight on the market. The Performance page showed the market forecasting better than the model alone, and a forecast that ignores that evidence would be less accurate on purpose. Matches without odds use the model alone, and the weight is stored with every prediction.
From the adjusted expected goals, a Poisson score matrix is computed — the probability of every scoreline from 0–0 upwards — and every football market is read off that one matrix: match result (1X2), over/under 2.5 goals, both teams to score, double chance, draw no bet, over/under 1.5 and 3.5 goals, each team scoring at least once, the home team scoring two or more, the correct score, and the half-time markets described below. Because they all come from the same forecast, they are always consistent with each other: over 1.5 is never less likely than over 2.5, and double chance is the sum of the two results it covers. For the correct score only the five most likely scorelines are shown, so those figures do not add up to 100%. Draw no bet is the home and away win probabilities rescaled to exclude the draw, since a drawn match returns the stake.
Half-time markets. The half-time result, over/under 0.5 and 1.5 first-half goals, and half-time/full-time are computed by splitting each team's expected goals between the two halves. The share scored before the break is measured for each league from the last three seasons — between 43% and 46% in the leagues covered, since more goals come late in matches — and cups use 45%. The two halves are combined into every half-time/full-time pairing, and those nine probabilities are then scaled so that their full-time totals match the match result forecast exactly.
For international fixtures (World Cup, European Championship), the same model is trained on national-team results going back several years, including competitive matches and friendlies. Friendlies are kept deliberately, because cross-confederation matches are the only games that let the model compare the relative strength of teams from different continents. The home advantage factor is removed for matches played at neutral venues.
Football Tournaments — Group Stage plus Knockout Bracket
Tournament competitions such as the World Cup and European Championship are not simple league tables, so their outright predictions (who reaches the final, who wins it) use a dedicated two-stage simulator built on top of the Dixon-Coles match engine described above. Each run simulates the full path to the trophy:
- Group stage. The remaining group fixtures are simulated with the same Dixon-Coles match engine. Each run produces a complete final standing for every group, including the ranking of third-placed teams across all groups where the format (such as the 48-team World Cup) allows the best third-placed teams to advance.
- Knockout bracket. The qualifiers are placed into the official bracket and every knockout round is simulated match by match, with ties resolved as they would be after extra time and penalties, until a champion emerges. The bracket structure, including how third-placed qualifiers are assigned to fixtures, follows the official competition regulations.
Aggregating thousands of these full-tournament runs produces, for every team, the probability of advancing from the group, reaching the quarter-finals, semi-finals, and final, and winning the tournament outright. These are the columns shown on a tournament's outright page. Because each run is considerably heavier than a single league simulation, tournaments use 5,000 runs per update rather than 10,000.
Formula 1 — Plackett-Luce Model
F1 predictions use a Plackett-Luce ranking model. Every driver in a race has a strength, and the finishing order is built one place at a time: the winner is drawn in proportion to the strengths of everyone on the grid, second place from the drivers left, and so on. Winner, podium and top-6 probabilities all come from the same simulated finishing orders, so they are always consistent with each other and add up exactly: the winner probabilities sum to 100%, the podium probabilities to 300%, the top-6 probabilities to 600%.
A driver's strength has three parts: the driver, the car for that season, and a smaller adjustment for how that driver is going this season compared with his career. The car is rated per team and per season because a team can build a winning car one year and a midfield car the next; each season's rating starts from the previous one and moves with the results. Teams that changed name keep their history (AlphaTauri to RB, Alfa Romeo to Sauber to Audi). The season adjustment lets a driver who improves sharply — typically in his second year — be rated on how he is driving now rather than on his rookie season.
The model is trained on every Grand Prix since 2022, with recent races weighted more heavily than older ones. Retirements are left out of the finishing order and modelled separately: each team's retirement rate is estimated from its history, and in every simulated race each car can retire before the order is drawn.
Predictions are published in two windows. Before the weekend, the forecast uses the strengths alone. After qualifying, the grid position is added with a weight the model has learned from past races — how much starting further forward is actually worth. The qualifying order stands in for the grid, because penalties applied later are not known yet. On sprint weekends the sprint winner forecast uses the sprint qualifying order once it exists.
Blend with a prediction market. Where a public prediction market prices the race winner (and, on sprint weekends, the sprint winner), its prices are captured twice a day during the race weekend. Prices at the very bottom of the book are a minimum quote rather than a forecast, so they are raised to a small floor, and the rest is rescaled so the market's probabilities add up to 100%. The model's and the market's winning chances are then combined with equal weight, and the podium and top-6 probabilities are simulated from that combined estimate, so the market's view reaches every F1 market while all of them stay consistent with each other. A capture older than 48 hours, or one that prices fewer than ten drivers, is not used, and the forecast is then the model alone. The equal weight is provisional: price history for past races was not available to measure it, so the model-only and market figures are stored with every forecast and the weight will be set from those records once enough races have been run.
This version replaced the original model on 30 October 2026. In a walk-forward backtest — each race predicted using only the races before it — the post-qualifying forecast was about 30% more accurate for the winner and the podium in 2026 than the original model, which did not use the grid. The test also showed that the grid position alone carries most of the information: the model adds most over it in the top-6 market. Before the weekend the gain is smaller.
A forecast is only published before the session that settles it: the race markets before the start of the race, the sprint winner before the start of the sprint. Once that moment has passed, the last forecast made before it stays as the published one.
Markets. For every Grand Prix the site publishes the race winner, podium, top 6 and top 10 (the points positions), and for each team the chance that each driver finishes ahead of his team-mate. All of these are read off the same simulated races, so they are always consistent with each other. In the team-mate battle the order of the official classification decides, and retirements are ranked there by laps completed, so every pair has a winner. On sprint weekends the sprint winner is published too.
Qualifying winner. Pole position has its own model, of the same kind as the race model but trained on qualifying results since 2022 instead of race results — who is quick over one lap is not quite who is quick over a race distance. In a backtest it was clearly better at predicting the pole sitter than the race model applied to qualifying. It is published before qualifying starts and does not use market prices.
The fastest lap market was withdrawn in October 2026. It no longer awards a point, it is decided largely by late pit stops for fresh tyres, and the model for it was weak. Forecasts already made for it remain on the Performance page.
WRC — Surface-Adjusted Elo
WRC predictions use a surface-adjusted Elo rating system combined with Bradley-Terry probability modelling.
Each driver maintains separate Elo ratings for overall performance, gravel rallies, asphalt rallies, and snow rallies. Ratings are updated after every event using finishing order and opponent strength, with higher K-factors for newer drivers so the model adapts faster when limited historical data exists.
Rally win and podium probabilities are derived from the Elo ratings, while reliability is modelled separately using each driver's historical finish rate. Drivers with frequent retirements due to crashes or mechanical issues receive lower effective probabilities.
Key factors include surface-specific performance, recent form, historical results at each rally, and overall reliability.
Tennis — Elo Ratings for ATP Singles
Tennis predictions use a head-to-head Elo rating system. Every player carries a single overall rating; the probability that one beats the other is the standard logistic function of the difference between their ratings, on the conventional 400-point scale. A 100-point gap means roughly a 64% chance for the higher-rated player, a 200-point gap roughly 76%.
Ratings move after every completed match by an amount that depends on how surprising the result was and on how much the model already knows about each player. The K-factor — the size of that adjustment — starts at 32 for a newcomer and decays toward 20 over roughly the first 80 rated matches. A qualifier arriving from the Challenger circuit therefore moves quickly on his first results, while an established player's rating barely shifts on a single match.
Results count according to the level of the tournament. A Grand Slam result moves ratings by the full amount, a Masters 1000 result by 90% of it, an ATP 500 result by 75% and an ATP 250 result by half. Earlier studies found no benefit from this; a later one, run after a fault in the order in which the match history was replayed had been corrected, found a small but consistent improvement, confirmed on a season that was not used to choose the weights. It has been in use since November 2026.
Grand Slams are played over five sets. Over a longer match the better player wins more often, because one bad set matters less. For Grand Slam matches the rating comparison is first converted into a chance of winning a single set, and from that into a chance of winning a best-of-five match. The published figure goes three quarters of the way towards that five-set number: the ratings themselves were learned from a mixture of three- and five-set matches, and the full conversion was not reliably better than the partial one when tested on past Grand Slams. The same conversion is used when a Grand Slam draw is simulated for the tournament outright.
Ratings are continuous across seasons. A player arrives in January carrying what he built over previous years rather than starting level with a debutant. Between seasons his match count is halved, which raises his K-factor again: the model treats an established player returning from a winter break as temporarily more uncertain, so his first few results of the year move his rating faster than they would mid-season.
Everything is rebuilt from scratch on every run rather than accumulated. The entire stored match history is replayed in chronological order each night, so a day the data collection missed leaves no permanent distortion — the backfilled match simply takes its place in the sequence. Until November 2026 the replay was not in true date order (it took the year's Grand Slam matches first, then each lower level in turn); that was corrected, and every rating was rebuilt.
A match is predicted only while it has not started. If the forecast cannot be made before the first ball, none is made: a figure written once a match is under way is not a forecast, and the published pre-match one is kept rather than replaced.
Walkovers and defaults are excluded from both the ratings and the accuracy scoring: they are administrative outcomes, not sporting ones, and rating them would credit a player for an opponent's illness. A match decided by a walkover therefore has no result on the Performance page — its forecast is void, like a draw-no-bet stake on a drawn football match. Retirements are treated as genuine results and count normally.
Markets. For every match the site publishes the winner and the score in sets; for every tournament, each player's chance of reaching the quarter-finals, the semi-finals and the final, and of winning the title. Doubles and qualifying matches are excluded throughout.
Tennis — Tournament Outright
Once a draw is published, the full bracket is simulated 10,000 times. Every tie is decided by the same rating comparison used for individual matches, winners advance through the bracket exactly as they would in the tournament, and the proportion of runs each player wins becomes his probability of taking the title.
A bye advances a player without simulating anything. It is the absence of a match, not an easy one, and treating it as a match would invent a chance of losing that does not exist.
No outright is published until the draw is complete. Qualifying finishes the day before the main draw, so a bracket published earlier still has empty places. Filling them with an invented rating would not simply add noise: the first published figure is the one the model is scored against, so a guess there would become the benchmark for everything that follows. Waiting costs about a day out of a market that runs for one to two weeks.
The estimate is republished daily as the tournament progresses, and each update starts from the players still in the draw rather than re-running the original field. A player knocked out in the first round therefore disappears from the figures rather than lingering at a few percent all week. Only the first snapshot — the one made before a ball was struck — counts towards the accuracy measured on the Performance page; the later ones show how the picture changed, and scoring them would flatter the model, since a forecast made once two players are left is barely a forecast at all.
For the same reason, a tournament that Proknoz only starts following midway through gets no outright at all. There is no honest baseline to publish.
Reaching each round. The same 10,000 simulated brackets that give the title chances also record how far each player got, so the chances of reaching the quarter-finals, semi-finals and final come from the very same runs and are always consistent with the title figure: across the field they add up to eight, four and two players. They follow the same rule as the title — only the snapshot taken before the first match counts on the Performance page — and are scored there against the round each player actually reached in the draw.
Tennis — Score in sets
The score in sets is derived from the published match-winner figure, so the two can never disagree: the match figure is converted into a chance of winning a single set, and from that into the chance of each final score — 2–0, 2–1, 1–2, 0–2, or 3–0 to 0–3 at Grand Slams.
The simplest version treats every set as an independent coin with the same bias, and tested on past matches it was clearly wrong: matches are more one-sided than that. The favourite won in straight sets about six points more often than it predicted (eleven at Grand Slams), and deciding sets were correspondingly rarer. A player who is having a good day tends to have it for the whole match. The model therefore adds a "form of the day": in each match a player's chance per set is shifted up or down by a fixed amount for the whole match, with equal probability, and the starting point is recalculated so that the match-winner figure is unchanged. The size of the shift was fitted on the 2025 season and checked on 2026, separately for each format: 0.20 for best-of-three, where every score then falls within two points of what happened in both seasons, and 0.18 for best-of-five, where the fit is a little looser — within three points, on a much smaller number of Grand Slam matches.
The score in sets is settled only for completed matches. After a retirement or a walkover there is no final score, so the forecast is void.
A first-set winner market was built the same way and tested, but not published: its calibration moved by several points between the two seasons measured, and it adds little to the match forecast.
Tennis — Surface ratings, and why they are shown but not used
Each player also carries separate ratings for hard, clay and grass courts, updated alongside the overall rating and displayed on player and match pages. Indoor hard courts are treated as hard courts.
These surface ratings do not currently influence any prediction, and that is a deliberate decision based on measurement rather than an oversight. Two separate studies, on two different datasets, found that blending the surface rating into the prediction produced no improvement at all.
The second study also found the reason. A surface rating only begins to exist when a player first plays on that surface, and it then accumulates roughly a third as many matches as his overall rating — so it permanently lags behind. In the sample measured, around 74% of hard-court ratings and half of clay ratings sat below the player's own overall rating. Because a weaker player's rating already sits near the average, that lag pulled the stronger player down further than the weaker one, and the net effect was to compress the gap between them rather than adjust for surface. On clay it was destroying about a fifth of the signal while appearing to refine it.
Turning the blend off removed a systematic under-confidence in the published probabilities: before the change, the favourite won more often than the model said in every band of the rating scale. The surface ratings are still maintained and still shown, because they are informative to a reader and because a corrected version of the blend may earn its place in future.
Tennis — What the model does not know
A new player starts at the tour average of 1500 regardless of his ATP ranking. This was measured before being accepted: matches involving players with fewer than three rated results are about 9% of the sample, and even a perfect starting estimate for them could improve overall accuracy by roughly 0.001 — smaller than the measurement noise. The fix would cost data-collection budget that is better spent elsewhere.
Head-to-head history, recent form and the average strength of recent opponents are shown on match pages as context, but none of them feeds the prediction. Each was tested as an input; none improved the forecasts, because the rating already reflects who a player has beaten. Head-to-head records cover every meeting Proknoz has recorded since 2024, on all surfaces and on each surface separately.
The betting market is better informed. Measured on the matches for which the data source publishes bookmaker prices — about half of them — the market's forecasts, with the bookmaker margin removed, are clearly more accurate than the model's: the gap in Brier score is about 0.009, ten times larger than any adjustment to the model itself has achieved. The market knows about injuries, fitness and conditions that results alone cannot show.
Blend with the betting market. Since 4 November 2026, where the data source publishes pre-match prices for a match — they come with the tournament draw, for matches not yet played — the bookmaker margin is removed and the model's probability is combined with the market's, with 60% of the weight on the market. Matches without a price use the model alone. A price of 1.90 on both players, which the source shows when there is no real market, is ignored. Tested on past matches, any weight on the market improved the forecasts, and the more weight the better; 60% was chosen deliberately below that, because the prices available for testing may be closing prices, which know more than the pre-match prices used live. The model's figure, the market's figure and the weight are stored with every prediction, so the weight can be re-measured on the prices actually used. When the market moves the published figure, the explanation says so. The tournament outright uses the model alone.
Predictions cover ATP singles only. Doubles, WTA, Challenger and ITF events are not modelled, although Challenger results would be a natural way to give newcomers a realistic starting rating.
Cycling — Stage-Type Regression
Cycling predictions use a logistic regression model trained separately for each stage type: flat (sprint), mountain, time trial, and mixed/hilly. Treating stage types as distinct prediction problems reflects the reality that specialists consistently outperform generalists in their preferred terrain.
Each rider is represented by a feature vector combining specialty scores (derived from ProCyclingStats profile ratings), recent form, and race fatigue. The specialty scores measure a rider's ability in six dimensions: sprinting, climbing, time trialling, punching power, one-day racing, and GC ability. Stage-type win rate — the fraction of top-5 finishes in stages of the same type over the previous two years — is the most predictive single feature and ensures that sprinters dominate sprint stages and climbers dominate mountain stages regardless of overall prestige or UCI points.
Fatigue is modelled in two components. Intra-race fatigue accumulates kilometres and elevation within the current race and is normalised against a full Grand Tour (approximately 3,500 km and 50,000 m). Inter-race fatigue measures total kilometres raced across all events in the previous 28 days, capturing the accumulated load from a packed calendar. Both components influence both predicted performance and DNF probability.
At prediction time, raw model scores are converted to a proper probability distribution via softmax across the entire field, so probabilities sum to exactly 100% across all starters.
Top-10 probabilities are drawn from that same distribution rather than derived from it by formula. Ten thousand finishing orders are sampled, each one picking a winner in proportion to the rider weights, removing him, picking the next from the remainder, and so on; a rider's top-10 probability is the fraction of those runs in which he finished in the first ten. Sampling rather than approximating keeps the two markets consistent with each other, and keeps the published figures away from 0% and 100% — ten thousand runs that all agree are evidence, not certainty.
Probabilities are computed across the full startlist but only the thirty strongest candidates are published for each stage. The figures shown on a stage page therefore do not sum to 100%: the remainder belongs to the rest of the peloton, whose individual chances are too small to be worth listing.
Models are trained on four seasons of WorldTour results (2022–2025) covering Grand Tours, monuments, and major stage races. After each stage completes, the relevant stage-type model is partially retrained with the new result, keeping predictions fresh as form evolves through the season.
Data is sourced from ProCyclingStats via FlareSolverr (Cloudflare bypass), with a rotating proxy pool to ensure reliable access. Historical stage results, rider profiles, and specialty scores are cached locally to minimise scraping frequency.
Cycling — GC and Classification Predictions
General classification (GC) predictions for stage races use Monte Carlo simulation over the remaining stages. Starting from the current live GC standings, each simulation run samples a time gain or loss for every active rider on each remaining stage, using distributions calibrated by stage type and rider specialty. GC contenders (high GC score or within five minutes of the leader) receive tighter distributions; domestiques and non-GC riders receive wider distributions reflecting their inconsistency in the mountains.
DNF probability increases with accumulated fatigue and each rider's historical retirement rate, and is applied independently on every simulated stage. Riders who abandon in a simulation run are excluded from subsequent stages in that run.
Mountains and Points classification probabilities are approximated from the GC simulation results using specialty score weighting: climber score for KOM, sprint score for Points. White jersey (best young rider) probabilities are derived by renormalising GC win probabilities over under-23 riders only.
For one-day classics, no GC simulation is needed — the stage winner model is applied directly to the full startlist.
Outright Predictions — Monte Carlo Simulation
League winner and championship predictions use Monte Carlo simulation. Remaining fixtures or events are simulated 10,000+ times using the current model parameters, producing a distribution of final standings.
In F1, each simulated race uses the same strengths and retirement rates as the Grand Prix forecasts, awards race and sprint points (there has been no point for the fastest lap since 2025), and only the drivers currently on the grid race the remaining rounds; a driver who has left keeps the points he scored. Cancelled rounds are not simulated.
In WRC, each simulated rally includes overall classification, Super Sunday, and Power Stage points. Manufacturer simulations also apply FIA scoring restrictions such as limiting points to nominated factory entries.
These probabilities are updated after every matchday, race weekend, rally, or stage so users can track how predictions evolve throughout the season.
Reasoning Explanations
Each prediction includes an automated reasoning summary. This is generated from templates — the model identifies the 3–5 most influential factors for that specific prediction, maps them to human-readable sentences, and composes a brief narrative. No large language models are involved in this step.
Depending on the sport, explanations may reference recent form, surface-specific performance, injuries and suspensions, qualifying pace, reliability, stage type suitability, accumulated fatigue, expected goals quality, confirmed lineup formations, or historical performance at a venue or rally.
For tennis, the explanation describes the size of the rating gap between the two players, and may reference their head-to-head record, recent results, and form on the match surface. The wording of the gap is calibrated against what actually happens: across the matches measured, a gap the model calls decisive was won by the favourite about 84% of the time, one called clear about 71%, and one called slight about 59%. Below that the match is described as close, and the favourite won about 54%. Head-to-head and surface form appear in the explanation as context for the reader — they are not inputs to the probability.
Tournament outright explanations are deliberately narrower. The bracket simulation reads two things and no more: each player's rating, and the shape of the draw. So the explanation reports where a player stands on rating among the field, and how often the simulation takes him to the final and lets him win it — and says nothing about his seeding, his recent form or his head-to-head record. Those are true facts about a player, but none of them was consulted, and presenting an unread number as a reason for the published one would be misleading however accurate the sentence.
Calibration and Accuracy
The Performance page tracks every published prediction against what actually happened. Three things about how it counts are worth stating plainly, because each of them changes the numbers.
The unit is a decision, not a line of probability. Predicting the winner of a Tour de France stage means assigning a probability to every rider in the peloton, and almost all of those lines are easy calls that the rider in question will not win. Counting each one as a separate prediction would let a single stage outweigh fifty football matches. One stage, one match, one title race counts once, and the score for it is the average across its lines.
Raw scores are not comparable between markets. A Brier score depends on how hard the question is. Guessing blindly scores about 0.25 on a tennis match, where there are two outcomes, and about 0.006 on a stage with 160 starters, where almost every line is correctly near zero. A low number in a large field is therefore not evidence of a good model. The Performance page shows a skill score alongside the raw Brier: how much better the model does than a forecaster who knows only how often this kind of call comes in. Zero means it adds nothing, one is perfect, and a negative number means it is worse than that baseline. The baseline itself is measured from what has happened, not assumed, and is computed separately for each market and each field size — winning a six-rider classic and winning a Grand Tour stage are the same market and not the same task.
Derived football markets are scored but not aggregated. Double chance, draw no bet, the other goal lines, team goals, correct score and the half-time markets are each scored and shown in their own row, but they are left out of the overall skill score, the sport summary, the monthly trend and the model version comparison. They restate the same scoreline forecast as the match result and over/under 2.5, so counting them as separate forecasts would multiply the weight of football in every total without adding information. A draw no bet forecast is not scored when the match is drawn, because the bet is void.
Derived Formula 1 markets are treated the same way. The top 10 and the team-mate battles come from the same simulated races as the winner, podium and top 6, so they are scored in their own rows but left out of the totals. The qualifying winner comes from a separate model and is counted.
And so are derived tennis markets. The score in sets restates the match-winner forecast, and the chances of reaching each round come from the same bracket simulation as the title, so both are scored in their own rows and left out of the totals. A score-in-sets forecast is not scored when the match ends in a retirement or a walkover.
Some predictions are deliberately not counted, and in every case it is because scoring them would flatter the model rather than measure it:
- A prediction written after the event started is not a forecast. Such rows exist in the database — a fixture that stays listed while a tournament runs can be re-predicted days after it was played — and they are excluded from scoring rather than deleted, so they remain available for diagnosis. The F1 sprint winner is held to the start of the sprint, not the start of the race on Sunday, because the sprint is decided the day before.
- An outright whose first snapshot was taken after the season ended is excluded for the same reason, as is one whose first snapshot named a single candidate. Recording a champion that was already decided is bookkeeping, not prediction, and it would score close to perfect.
- A skill score is not published where the baseline or the sample is too thin to support it. A figure built on two races cannot carry three decimal places, and printing it in the same column as one built on a hundred and fifty would invite the reader to compare them.
Football is also measured against the betting market. For match result and over/under 2.5 goals, the Performance page compares the model with the market's closing odds — the average price across bookmakers just before kick-off, which reflects everything the market knew about team news and form. Tennis match winner forecasts are compared in the same way with the two prices the tennis data source publishes for each match, where it publishes them; that source does not say when the prices were taken, so the tennis column is a comparison with the market as recorded rather than strictly with the closing line. The bookmaker's margin is removed by scaling the implied probabilities to sum to 100%, and both forecasts are scored on exactly the same matches. The comparison is only published once at least thirty priced matches exist in a market. Since 29 October 2026 the forecasts also blend in the market's earlier pre-match prices (see the football model above), so the comparison now measures whether the published forecast improves on the closing price. The closing odds themselves are never an input: they are only known at kick-off, after the forecast has been published.
Every input the football model used for a prediction — the absences it counted, the xG figures and the size of each adjustment — is stored with the prediction at the moment it is made. That record is what allows the size of those adjustments to be tested against results later, rather than set by judgement alone.
Because the baseline is measured rather than assumed, the page can and does report that a model is worse than guessing. That is the intended behaviour. A tracking page that can only produce flattering numbers is not tracking anything.
Data Sources
All data comes from free, publicly available sources.
For football: Football-Data.co.uk (historical CSV data for model training), Football-Data.org (fixtures and results API), the martj42 international results dataset (national-team match history for World Cup and Euro modelling), Understat (per-match and season xG for the top five European leagues, back to 2021), SofaScore (per-match xG for Liga Portugal and the Championship), and Transfermarkt (player injuries, suspensions, and confirmed lineups — scraped from public league pages). Football-Data.co.uk also supplies the market-average pre-match odds blended into the forecasts, and the closing odds used to benchmark them on the Performance page.
For Formula 1: the FastF1 library, which provides qualifying times, lap data, pit stops, and tyre information from the official F1 Live Timing feed, and the Jolpica API (the successor of the Ergast database), which provides the official race and sprint classifications, points, grid positions, and the race history since 2022 used to train the model. Race and sprint winner prices come from a public prediction market.
For WRC: the official WRC.com public API, which provides structured event data including itineraries, entry lists, stage results, rally classifications, and championship standings.
For cycling: ProCyclingStats, which provides race calendars, stage profiles, startlists, results, and rider specialty profiles for all WorldTour events.
For tennis: the tennis-api (Matchstat) service, which provides the ATP calendar, tournament results, match details, the draw bracket for each tournament, and bookmaker prices: the pre-match prices that come with the draw are blended into the forecasts, and the prices recorded with each result are used to benchmark them. Head-to-head records are not requested from the source — they are derived from the matches already stored, which covers every meeting Proknoz has recorded.
Limitations
Models are simplifications of reality. They don't fully account for tactical changes, transfer-window signings, managerial motivation, weather conditions beyond basic factors, or unpredictable events.
Football injury data from Transfermarkt reflects publicly reported absences, which may lag behind actual squad news by a day or two. Players listed as doubtful are corrected only by the lineup re-prediction about an hour before kick-off, and only when both lineups and enough lineup history for each team are available; otherwise the injury list stands. The market blend is only as current as the odds Football-Data.co.uk has published, which may be several days old. The xG adjustment applies season-long averages and does not capture recent shooting form changes within the season. European cup competitions have no xG source and receive no xG adjustment.
For international football, comparing the strength of teams from different confederations is inherently difficult, because they rarely play each other outside major tournaments. The model leans on friendlies and intercontinental fixtures to calibrate this, and applies regularisation to limit over-rating, but some teams that dominate weaker regional opposition may still be rated more highly than a neutral observer would expect. Tournament outright probabilities should be read in that light, particularly in the early rounds when few matches have been played. As more matches are played and tracked on the Performance page, these estimates can be assessed against actual outcomes.
Tournament outright simulations currently treat all knockout matches as played at neutral venues and do not model travel, altitude, host-nation advantage, or squad rotation between fixtures. Extra time and penalty shoot-outs are modelled as a single tie-break rather than as separately simulated phases.
F1 predictions do not model weather, safety cars, team orders, tyre strategy, or circuit-specific characteristics: a driver's strength is the same at Monaco and at Monza. Grid penalties are not known when the post-qualifying forecast is made. The first season of new technical regulations is harder to predict, because the cars' order can change completely from one year to the next and the model needs a few races to see it.
WRC predictions do not currently model road position effects on gravel, changing weather conditions during rallies, or team orders.
Tennis predictions know nothing about injuries, fitness, travel, altitude, court speed, weather, or a player's motivation for a particular event — all of which matter in a sport decided by one athlete against another. A rating summarises results and nothing else. Ratings for players new to the tour are unreliable until they have accumulated a season or so of results.
The outright simulation inherits every one of those blind spots and compounds them, because a tournament is six or seven matches and the errors accumulate along the path. It also assumes the draw is played as published: a withdrawal after the bracket is set is not reflected until the following day's update. And it does not distinguish a favourable draw from an unfavourable one in the explanation, because measuring that properly requires comparing the actual bracket against an average one, which is not yet done.
Cycling predictions do not model team tactics (leadout trains, domestique support, protected riders), breakaway probability, or intermediate sprint bonuses in the GC simulation. One-day race predictions treat each event independently and do not account for fatigue from earlier races in the same week (e.g. the Ardennes classics).
Cycling stage predictions rank a field the model has seen only through results and profile ratings. A rider in unusual form, a team riding for a team-mate, or a breakaway that is allowed to stay away will all beat the model, and the winner of a stage is frequently outside the thirty riders it considered most likely. The Performance page measures how often that happens rather than hiding it.
Probability estimates are the model's best assessment, not certainty.
The models are trained primarily on recent seasons and may perform less well for newly promoted teams, rookie riders, or competitions with limited historical data.