Calibration & track record
Football Prediction Accuracy: How Accurate Is TACTIX?
We answer with calibration: when the model gives an outcome a 60% chance, that kind of call should come true about 60% of the time. Here is how close it gets, and how it is measured.
Football prediction accuracy at TACTIX is measured as calibration, and across 6M+ graded predictions (all markets, player markets included) the model sits within ±0.2% of how often those outcomes actually happened, counting each prediction graded since the public record began. Winners and losers both count, in every market the model prices, and the figure is recomputed from the database on a schedule rather than typed in.
The Calibration Chart, With Every Sample Size
Each circle is one 10-point probability bucket: how likely the model said its outcomes were, against how often they actually happened. A perfectly calibrated model sits on the dashed line. Winners and losers are both in every bucket, and the table gives the number of predictions behind each point.
- Dot area = settled predictions in the band (smallest 450,286, largest 1,479,060)
- Whisker = 95% Wilson interval
- Dashed line = perfect calibration
- Gold, Silver and Bronze mark each tier's hit rate at its confidence floor
Source: TACTIX settled predictions (all markets) · as of 16 Nov 2026, 19:00 UK · n = 6,747,459 settled predictions
Show the Numbers
| 4% | 4% | 1,479,060 | 4.4–4.4% |
|---|---|---|---|
| 15% | 15% | 713,118 | 14.6–14.8% |
| 25% | 25% | 553,666 | 24.8–25% |
| 35% | 36% | 489,436 | 35.4–35.6% |
| 45% | 45% | 510,183 | 45.2–45.4% |
| 55% | 55% | 472,193 | 54.6–54.8% |
| 65% | 65% | 450,286 | 64.4–64.6% |
| 75% | 75% | 483,098 | 74.9–75.1% |
| 85% | 85% | 583,731 | 85.3–85.5% |
| 96% | 95% | 1,012,688 | 95.4–95.4% |
| Bronze (70%+) | 87.8% | 1,596,686 | 87.7–87.9% |
| Silver (80%+) | 91.7% | 1,219,841 | 91.7–91.8% |
| Gold (90%+) | 95.4% | 777,229 | 95.3–95.5% |
| 0–10% | 4.3% | 4.4% | 1,479,060 |
| 10–20% | 14.6% | 14.7% | 713,118 |
| 20–30% | 24.9% | 24.9% | 553,666 |
| 30–40% | 35.1% | 35.5% | 489,436 |
| 40–50% | 45.0% | 45.3% | 510,183 |
| 50–60% | 54.9% | 54.7% | 472,193 |
| 60–70% | 64.9% | 64.5% | 450,286 |
| 70–80% | 75.0% | 75.0% | 483,098 |
| 80–90% | 85.4% | 85.4% | 583,731 |
| 90–100% | 95.7% | 95.4% | 1,012,688 |
How to Read the Reliability Curve
Each dot pools the graded predictions made at a given confidence level. Its position answers one question: when the model said "X% likely", how often did that outcome actually happen? A dot at x = 0.70, y = 0.69 means outcomes the model priced at 70% occurred 69% of the time. The dashed diagonal is perfection; the further a dot sits from it, the more the model over- or under-states that probability band. The headline figure — ±0.2% across 6M+ settled predictions (all markets, player markets included) — is the average of those gaps, weighted by sample.
Terms are defined in the glossary; the method itself is on the methodology page.
Hit Rate by Confidence Tier
- Gold90%+ confidence95.4%741,271 of 777,229 picks from 7,221 matches (95–95%)
- Silver80%+ confidence91.7%1,119,196 of 1,219,841 picks from 7,222 matches (92–92%)
- Bronze70%+ confidence87.8%1,401,769 of 1,596,686 picks from 7,222 matches (88–88%)
Ranges in brackets are 95% intervals on the rated picks' hit rate, counting each match once: picks in the same match tend to land or miss together.
The graded predictions, all markets, each rate shown with the number it is measured on. The tier is the model’s own confidence band — the threshold beside each one is where it starts.
As of .
What the Confidence Tiers Mean
Every pick carries a model probability. We bucket them into three fixed-threshold tiers and report each one’s real win rate. The tiers are fixed probability thresholds, so a “Gold” pick means the same thing in every market and league.
- Gold — highest-confidence calls: 95% hit rate, 741,271 of 777,229 picks from 7,221 matches (95–95%).
- Silver — strong leans: 92% hit rate, 1,119,196 of 1,219,841 picks from 7,222 matches (92–92%).
- Bronze — solid value: 88% hit rate, 1,401,769 of 1,596,686 picks from 7,222 matches (88–88%).
Ranges in brackets are 95% intervals on the rated picks' hit rate, counting each match once: picks in the same match tend to land or miss together.
These tier rates are pooled across every market, so the headline average leans towards high-volume markets like over/under goals. Match result (1X2) — the hardest market to call — is shown on its own row in the table below; read the by-market figures, not the pooled average, as the honest measure of accuracy.
Accuracy by Market
| Total goals (over / under) | 93% 10,352 of 11,143 picks from 7,129 matches (92–93%) | 89% 19,190 of 21,605 picks from 7,130 matches (88–89%) | 85% 25,126 of 29,523 picks from 7,130 matches (85–86%) | 62,271 |
| Double chance | — | 76% 329 of 433 picks from 343 matches (72–80%) | 75% 7,209 of 9,604 picks from 6,435 matches (74–76%) | 10,086 |
| Team to score | — | 83% 1,800 of 2,174 picks from 2,005 matches (81–84%) | 77% 7,707 of 10,028 picks from 6,415 matches (76–78%) | 12,285 |
| First-half goals | 90% 517 of 577 picks from 570 matches (87–92%) | 85% 5,147 of 6,041 picks from 5,862 matches (84–86%) | 80% 9,611 of 12,006 picks from 7,050 matches (79–81%) | 18,624 |
| Both teams to score | — | 56% 1,326 of 2,388 picks from 2,388 matches (54–58%) | 56% 9,445 of 16,965 picks from 16,965 matches (55–56%) | 19,530 |
Track Record by League & Market
Calibration error for every league and market we cover — losers included, no cherry-picking. Lower is better: ±2% means that when we say 60%, it lands about 58–62% of the time. Filter by league, market and window.
Showing 25 of 16024 cells · all-time · model rule-based-v2 · updated . Lower calibration error = more honest probabilities; cells under 200 settled predictions are flagged “low sample”.
| Friendlies | Match result (1X2) | 1k | ±2.4% | Good |
| FA Cup (England) | Match result (1X2) | 1k | ±4.0% | Fair |
| UEFA Europa Conference League | Match result (1X2) | 939 | ±2.7% | Good |
| Major League Soccer | Match result (1X2) | 927 | ±5.2% | Fair |
| Liga Profesional Argentina | Match result (1X2) | 726 | ±3.5% | Good |
| National League | Match result (1X2) | 600 | ±1.6% | Excellent |
| Serie A (Brazil) | Match result (1X2) | 570 | ±3.4% | Good |
| Championship (England) | Match result (1X2) | 543 | ±3.8% | Good |
| Segunda División | Match result (1X2) | 543 | ±5.4% | Fair |
| League Two (England) | Match result (1X2) | 525 | ±6.0% | Variable |
| NWSL Women | Match result (1X2) | 519 | ±6.4% | Variable |
| Allsvenskan | Match result (1X2) | 495 | ±3.4% | Good |
| League One (England) | Match result (1X2) | 492 | ±7.9% | Variable |
| UEFA Champions League | Match result (1X2) | 489 | ±3.2% | Good |
| UEFA Europa League | Match result (1X2) | 459 | ±3.0% | Good |
| Super League (China) | Match result (1X2) | 432 | ±3.7% | Good |
| Liga MX | Match result (1X2) | 432 | ±9.0% | Variable |
| La Liga | Match result (1X2) | 417 | ±5.7% | Fair |
| UEFA Nations League | Match result (1X2) | 414 | ±4.4% | Fair |
| Eliteserien | Match result (1X2) | 411 | ±3.8% | Good |
| Ekstraklasa | Match result (1X2) | 408 | ±3.1% | Good |
| Challenge Cup | Match result (1X2) | 408 | ±5.2% | Fair |
| Serie A (Italy) | Match result (1X2) | 390 | ±5.8% | Fair |
| Premier League (Wales) | Match result (1X2) | 381 | ±6.7% | Variable |
| K League 1 | Match result (1X2) | 366 | ±10.7% | Variable |
The Latest Monthly Scorecard — October 2026
Every month we publish the same measurement for that month alone, good months and bad. In October 2026: 2,158,088 settled predictions, with stated and observed rates differing by 0.4 percentage points on average.
In October 2026 the bookmakers’ closing prices were more accurate than the model (log-loss difference +0.022, 95% interval +0.017 to +0.028 across 31 match days). The comparison uses the 11,158 of 38,471 predictions that had a closing price captured within 6 hours of kick-off.
Read the October 2026 scorecard · the latest scorecard, always
Against the Bookmakers’ Closing Prices
For match-result, over/under 2.5 goals and both-teams-to-score predictions, we compare each probability with the bookmakers’ closing price for the same outcome, with the margin removed. A closing price counts only when it was captured within 6 hours of kick-off. In October 2026 a closing price was available for 11,158 of 38,471 of those predictions.
In October 2026 the bookmakers’ closing prices were more accurate than the model (log-loss difference +0.022, 95% interval +0.017 to +0.028 across 31 match days). Across those 11,158 paired predictions the model’s Brier score was 0.222 and the closing prices’ was 0.212 (lower is better for both).
| Match result (1X2) | 5,442 of 6,789 | 0.197 | 0.186 | closing prices more accurate |
| Over/under 2.5 goals | 3,422 of 27,156 | 0.245 | 0.235 | closing prices more accurate |
| Both teams to score | 2,294 of 4,526 | 0.249 | 0.242 | closing prices more accurate |
On the 1,814 matches with a complete closing 1X2 price, the ranked probability score was 0.209 for the model and 0.192 for the closing prices (lower is better).
| 0–10% | 4.2% | 4.1% | 463,377 |
| 10–20% | 14.6% | 14.1% | 237,635 |
| 20–30% | 24.9% | 24.2% | 169,862 |
| 30–40% | 35.2% | 34.6% | 152,190 |
| 40–50% | 45.1% | 45.1% | 159,975 |
| 50–60% | 54.8% | 54.8% | 145,932 |
| 60–70% | 64.9% | 65.5% | 144,345 |
| 70–80% | 75.0% | 76.0% | 150,215 |
| 80–90% | 85.5% | 86.1% | 201,548 |
| 90–100% | 95.7% | 95.8% | 333,009 |
As of .
Verified Live — With the Losers Shown
Our calibration and track record update from settled results in public, losing predictions included. You can watch it on the live scoreboard. We never hide bad weeks and we never sell "guaranteed" tips — a prediction is a probability, not a certainty, and no model reliably beats the bookmakers' margin.
- NJ/NY Gotham FC W 1–2 Kansas City WHome or Away · 73.2% · BronzeLanded
- Eldense 2–0 SabadellHome or Away · 71.6% · BronzeLanded
- Eldense 2–0 SabadellOver 1.5 goals · 77.6% · BronzeLanded
- West Ham W 2–0 Charlton Athletic WHome win · 71.4% · BronzeLanded
- Sutton Utd 0–2 Solihull MoorsHome or Draw · 74% · BronzeMissed
- Kidderminster Harriers 1–1 FC Halifax TownHome or Away · 74% · BronzeMissed
- Crewe 2–2 Newport CountyHome or Away · 75% · BronzeMissed
- Colchester 2–0 GillinghamHome or Away · 76.1% · BronzeLanded
- Shrewsbury 2–1 OldhamOver 1.5 goals · 81% · SilverLanded
- Leganes 3–1 BurgosHome or Away · 73.7% · BronzeLanded
- Leganes 3–1 BurgosOver 1.5 goals · 79.4% · BronzeLanded
- SL Benfica W 3–1 Austria Wien WUnder 3.5 goals · 80.8% · SilverMissed
Do Hot Streaks Last?
Each observation is one appearance by one player for one market line. Appearances are taken oldest first and counted as in the hit meters (last eight played matches, friendlies left out, appearances of 15+ minutes); a match where the statistic was not recorded for the player’s team is left out, and a match the player did not appear in is not counted at all. For every appearance with at least eight counted appearances before it, we note how many of those eight cleared the line (zero to eight) and whether this one did. Intervals are Wilson 95% intervals. A row with fewer than 200 observations shows its count only.
- Shots (line 1.5): players who had hit the line in eight of their last eight hit it in their next appearance 2,444 of 3,432 times (71.2%), against 17.7% for all 1,349,151 observations. The intervals do not overlap.
- Shots on target (line 0.5): players who had hit the line in eight of their last eight hit it in their next appearance 2,999 of 4,475 times (67.0%), against 24.9% for all 1,307,571 observations. The intervals do not overlap.
- Goal or assist (line 0.5): players who had hit the line in eight of their last eight hit it in their next appearance 173 of 300 times (57.7%), against 14.6% for all 1,352,011 observations. The intervals do not overlap.
- Fouls committed (line 1.5): players who had hit the line in eight of their last eight hit it in their next appearance 499 of 880 times (56.7%), against 22.5% for all 1,349,913 observations. The intervals do not overlap.
- Fouls won (line 0.5): players who had hit the line in eight of their last eight hit it in their next appearance 38,607 of 50,081 times (77.1%), against 50.7% for all 1,349,875 observations. The intervals do not overlap.
- Tackles (line 1.5): players who had hit the line in eight of their last eight hit it in their next appearance 3,720 of 5,771 times (64.5%), against 30.7% for all 1,304,232 observations. The intervals do not overlap.
- Offsides (line 0.5): players who had hit the line in eight of their last eight hit it in their next appearance 463 of 767 times (60.4%), against 13.3% for all 872,944 observations. The intervals do not overlap.
- Saves (goalkeepers) (line 2.5): players who had hit the line in eight of their last eight hit it in their next appearance 970 of 1,509 times (64.3%), against 55.4% for all 93,731 observations. The intervals do not overlap.
| 0 of 8 | 545,792 | 33,791 (6.2%) | 6.1% to 6.3% |
|---|---|---|---|
| 1 of 8 | 307,770 | 43,453 (14.1%) | 14.0% to 14.2% |
| 2 of 8 | 194,938 | 43,805 (22.5%) | 22.3% to 22.7% |
| 3 of 8 | 126,977 | 38,956 (30.7%) | 30.4% to 30.9% |
| 4 of 8 | 82,108 | 31,800 (38.7%) | 38.4% to 39.1% |
| 5 of 8 | 49,834 | 23,153 (46.5%) | 46.0% to 46.9% |
| 6 of 8 | 26,595 | 14,266 (53.6%) | 53.0% to 54.2% |
| 7 of 8 | 11,705 | 7,292 (62.3%) | 61.4% to 63.2% |
| 8 of 8 | 3,432 | 2,444 (71.2%) | 69.7% to 72.7% |
| All observations | 1,349,151 | 238,960 (17.7%) | 17.7% to 17.8% |
| 0 of 8 | 337,921 | 29,893 (8.9%) | 8.8% to 8.9% |
|---|---|---|---|
| 1 of 8 | 286,241 | 50,400 (17.6%) | 17.5% to 17.8% |
| 2 of 8 | 234,358 | 57,702 (24.6%) | 24.5% to 24.8% |
| 3 of 8 | 174,304 | 57,697 (33.1%) | 32.9% to 33.3% |
| 4 of 8 | 124,241 | 51,167 (41.2%) | 40.9% to 41.5% |
| 5 of 8 | 82,238 | 39,635 (48.2%) | 47.9% to 48.5% |
| 6 of 8 | 45,088 | 24,709 (54.8%) | 54.3% to 55.3% |
| 7 of 8 | 18,705 | 11,404 (61.0%) | 60.3% to 61.7% |
| 8 of 8 | 4,475 | 2,999 (67.0%) | 65.6% to 68.4% |
| All observations | 1,307,571 | 325,606 (24.9%) | 24.8% to 25.0% |
| 0 of 8 | 768,012 | 37,439 (4.9%) | 4.8% to 4.9% |
|---|---|---|---|
| 1 of 8 | 336,528 | 34,654 (10.3%) | 10.2% to 10.4% |
| 2 of 8 | 146,750 | 23,941 (16.3%) | 16.1% to 16.5% |
| 3 of 8 | 62,929 | 14,217 (22.6%) | 22.3% to 22.9% |
| 4 of 8 | 25,664 | 7,282 (28.4%) | 27.8% to 28.9% |
| 5 of 8 | 9,090 | 3,009 (33.1%) | 32.1% to 34.1% |
| 6 of 8 | 2,427 | 922 (38.0%) | 36.1% to 39.9% |
| 7 of 8 | 528 | 260 (49.2%) | 45.0% to 53.5% |
| 8 of 8 | 83 | count only | – |
| All observations | 1,352,011 | 121,772 (9.0%) | 9.0% to 9.1% |
| 0 of 8 | 839,636 | 41,091 (4.9%) | 4.9% to 4.9% |
|---|---|---|---|
| 1 of 8 | 357,764 | 29,066 (8.1%) | 8.0% to 8.2% |
| 2 of 8 | 115,905 | 12,991 (11.2%) | 11.0% to 11.4% |
| 3 of 8 | 30,546 | 4,322 (14.2%) | 13.8% to 14.5% |
| 4 of 8 | 6,738 | 1,186 (17.6%) | 16.7% to 18.5% |
| 5 of 8 | 1,215 | 265 (21.8%) | 19.6% to 24.2% |
| 6 of 8 | 176 | count only | – |
| 7 of 8 | 31 | count only | – |
| 8 of 8 | 0 | count only | – |
| All observations | 1,352,011 | 88,980 (6.6%) | 6.5% to 6.6% |
| 0 of 8 | 537,322 | 40,782 (7.6%) | 7.5% to 7.7% |
|---|---|---|---|
| 1 of 8 | 379,399 | 51,517 (13.6%) | 13.5% to 13.7% |
| 2 of 8 | 225,544 | 43,980 (19.5%) | 19.3% to 19.7% |
| 3 of 8 | 120,087 | 30,695 (25.6%) | 25.3% to 25.8% |
| 4 of 8 | 56,605 | 17,663 (31.2%) | 30.8% to 31.6% |
| 5 of 8 | 23,401 | 8,584 (36.7%) | 36.1% to 37.3% |
| 6 of 8 | 7,509 | 3,159 (42.1%) | 41.0% to 43.2% |
| 7 of 8 | 1,844 | 941 (51.0%) | 48.8% to 53.3% |
| 8 of 8 | 300 | 173 (57.7%) | 52.0% to 63.1% |
| All observations | 1,352,011 | 197,494 (14.6%) | 14.6% to 14.7% |
| 0 of 8 | 322,028 | 32,613 (10.1%) | 10.0% to 10.2% |
|---|---|---|---|
| 1 of 8 | 323,717 | 63,312 (19.6%) | 19.4% to 19.7% |
| 2 of 8 | 295,216 | 71,852 (24.3%) | 24.2% to 24.5% |
| 3 of 8 | 207,487 | 60,406 (29.1%) | 28.9% to 29.3% |
| 4 of 8 | 118,763 | 40,751 (34.3%) | 34.0% to 34.6% |
| 5 of 8 | 55,442 | 21,800 (39.3%) | 38.9% to 39.7% |
| 6 of 8 | 20,738 | 9,280 (44.8%) | 44.1% to 45.4% |
| 7 of 8 | 5,642 | 2,852 (50.5%) | 49.2% to 51.9% |
| 8 of 8 | 880 | 499 (56.7%) | 53.4% to 59.9% |
| All observations | 1,349,913 | 303,365 (22.5%) | 22.4% to 22.5% |
| 0 of 8 | 53,019 | 9,907 (18.7%) | 18.4% to 19.0% |
|---|---|---|---|
| 1 of 8 | 107,511 | 29,163 (27.1%) | 26.9% to 27.4% |
| 2 of 8 | 158,641 | 57,331 (36.1%) | 35.9% to 36.4% |
| 3 of 8 | 206,405 | 91,065 (44.1%) | 43.9% to 44.3% |
| 4 of 8 | 233,541 | 119,596 (51.2%) | 51.0% to 51.4% |
| 5 of 8 | 229,278 | 131,625 (57.4%) | 57.2% to 57.6% |
| 6 of 8 | 188,437 | 120,767 (64.1%) | 63.9% to 64.3% |
| 7 of 8 | 122,962 | 86,567 (70.4%) | 70.2% to 70.7% |
| 8 of 8 | 50,081 | 38,607 (77.1%) | 76.7% to 77.5% |
| All observations | 1,349,875 | 684,628 (50.7%) | 50.6% to 50.8% |
| 0 of 8 | 255,939 | 21,461 (8.4%) | 8.3% to 8.5% |
|---|---|---|---|
| 1 of 8 | 214,766 | 46,976 (21.9%) | 21.7% to 22.1% |
| 2 of 8 | 230,705 | 68,426 (29.7%) | 29.5% to 29.9% |
| 3 of 8 | 216,346 | 79,134 (36.6%) | 36.4% to 36.8% |
| 4 of 8 | 174,118 | 73,879 (42.4%) | 42.2% to 42.7% |
| 5 of 8 | 117,587 | 56,637 (48.2%) | 47.9% to 48.5% |
| 6 of 8 | 63,265 | 33,980 (53.7%) | 53.3% to 54.1% |
| 7 of 8 | 25,735 | 15,589 (60.6%) | 60.0% to 61.2% |
| 8 of 8 | 5,771 | 3,720 (64.5%) | 63.2% to 65.7% |
| All observations | 1,304,232 | 399,802 (30.7%) | 30.6% to 30.7% |
| 0 of 8 | 465,174 | 49,275 (10.6%) | 10.5% to 10.7% |
|---|---|---|---|
| 1 of 8 | 462,644 | 63,903 (13.8%) | 13.7% to 13.9% |
| 2 of 8 | 270,590 | 45,219 (16.7%) | 16.6% to 16.9% |
| 3 of 8 | 112,189 | 21,335 (19.0%) | 18.8% to 19.3% |
| 4 of 8 | 33,182 | 7,024 (21.2%) | 20.7% to 21.6% |
| 5 of 8 | 7,095 | 1,617 (22.8%) | 21.8% to 23.8% |
| 6 of 8 | 1,018 | 253 (24.9%) | 22.3% to 27.6% |
| 7 of 8 | 112 | count only | – |
| 8 of 8 | 7 | count only | – |
| All observations | 1,352,011 | 188,664 (14.0%) | 13.9% to 14.0% |
| 0 of 8 | 445,047 | 21,751 (4.9%) | 4.8% to 5.0% |
|---|---|---|---|
| 1 of 8 | 193,204 | 22,790 (11.8%) | 11.7% to 11.9% |
| 2 of 8 | 100,326 | 20,886 (20.8%) | 20.6% to 21.1% |
| 3 of 8 | 60,049 | 18,132 (30.2%) | 29.8% to 30.6% |
| 4 of 8 | 37,421 | 14,309 (38.2%) | 37.8% to 38.7% |
| 5 of 8 | 21,803 | 9,824 (45.1%) | 44.4% to 45.7% |
| 6 of 8 | 10,546 | 5,414 (51.3%) | 50.4% to 52.3% |
| 7 of 8 | 3,781 | 2,106 (55.7%) | 54.1% to 57.3% |
| 8 of 8 | 767 | 463 (60.4%) | 56.9% to 63.8% |
| All observations | 872,944 | 115,675 (13.3%) | 13.2% to 13.3% |
| 0 of 8 | 1,041 | 142 (13.6%) | 11.7% to 15.9% |
|---|---|---|---|
| 1 of 8 | 2,115 | 998 (47.2%) | 45.1% to 49.3% |
| 2 of 8 | 7,106 | 3,614 (50.9%) | 49.7% to 52.0% |
| 3 of 8 | 15,181 | 7,917 (52.2%) | 51.4% to 52.9% |
| 4 of 8 | 21,973 | 12,022 (54.7%) | 54.1% to 55.4% |
| 5 of 8 | 22,268 | 12,733 (57.2%) | 56.5% to 57.8% |
| 6 of 8 | 15,583 | 9,186 (59.0%) | 58.2% to 59.7% |
| 7 of 8 | 6,955 | 4,295 (61.8%) | 60.6% to 62.9% |
| 8 of 8 | 1,509 | 970 (64.3%) | 61.8% to 66.7% |
| All observations | 93,731 | 51,877 (55.4%) | 55.0% to 55.7% |
Counts of past appearances. Not a forecast. As of 16 November 2026, from 13,035,461 observations. The cells as CSV · the same figures as JSON. 18+ | BeGambleAware.org
Why Is Calibration the Right Test of Football Prediction Accuracy?
A single "accuracy %" mostly reflects how predictable a league or market is, and it is easy to flatter by picking easy markets. Calibration asks whether the stated probabilities are honest: when the model says 70%, does that outcome happen about seven times in ten? That is the property that matters when you use a probability to make a decision, so it is the number we publish, next to the count of predictions behind it.
How We Measure It
- Every graded prediction counts — winners and losers alike, across every market the model prices. Nothing is removed for looking bad.Predictions that could not be graded against a result (n = 2,426,343, all-time, all markets, player markets included, counted 2026-11-16) are marked void and are not in these figures. Each match is graded on the forecasts of the one model that served it. Pre-match forecasts only; 4,257,582 written after kick-off are not counted. Since 1 Nov 2026, the record counts forecasts written before kick-off only.
- Calibration error is the average gap, in percentage points, between the probability the model stated and how often those outcomes happened, measured in 10-point buckets and weighted by the number of predictions in each.
- The figure is recomputed from the database every six hours, and the page shows the result of the latest run with the number of predictions it covers.
- Each month is published on its own: a monthly scorecard counts only the matches that kicked off in that month, and compares the model with the bookmakers’ closing prices, margin removed — whichever way that comparison falls.
How the Model Is Built
Every TACTIX forecast comes from one published model, calibrated market-by-market. A dedicated three-way match-result model blends expected goals (xG) with a team Elo rating. On top of its base model, the current model adds per-market calibration (each market tuned to its own reliability), a dedicated three-way match-result model that combines expected goals with team Elo ratings, and, where it is switched on, a stacking layer over the base model's probabilities. Full detail is on our methodology page and the live performance is on the Model Scoreboard.
Frequently Asked
- How accurate are TACTIX's football predictions?
- TACTIX's model is calibrated to within ±0.2% across 6M+ settled predictions (all markets, player markets included). When it states a 60% chance, the outcome happens close to 60% of the time. We publish calibration error rather than a single accuracy figure because it is the honest measure of a probability.
- How often do TACTIX's Gold, Silver and Bronze picks land?
- By confidence tier, settled to date: Gold (a model probability of 90% or more): 95% hit rate, 741,271 of 777,229 picks from 7,221 matches (95–95%); Silver (80% or more, Gold included): 92% hit rate, 1,119,196 of 1,219,841 picks from 7,222 matches (92–92%); Bronze (70% or more, Silver and Gold included): 88% hit rate, 1,401,769 of 1,596,686 picks from 7,222 matches (88–88%). These rates are pooled across every market, so they lean towards high-volume markets such as over/under goals. Ranges in brackets are 95% intervals on the rated picks' hit rate, counting each match once: picks in the same match tend to land or miss together.
- What does "calibrated" mean?
- A calibrated model's stated probabilities match reality over the long run: events it calls 30% happen about 30% of the time, 70% about 70%. It is the property that makes a probability trustworthy for decisions.
- Does TACTIX guarantee winning bets?
- No. A prediction is a probability, not a certainty. Football is high-variance and no model reliably beats the bookmakers' margin. TACTIX publishes calibrated probabilities and its track record, losers included, and never sells guaranteed tips.
- How is the TACTIX model built?
- It is one published model, calibrated market-by-market, with a three-way match-result model that blends expected goals (xG) and a team Elo rating.
- Can I see TACTIX's track record?
- Yes — live on the scoreboard, with losing predictions included, so the record is verifiable rather than marketing.
- How do I get TACTIX's predictions?
- Each match page carries a free model pick. The full win probability, expected-goals breakdown and value across 240+ markets are part of TACTIX Pro.
Responsible Gambling
Anywhere we discuss betting we carry 18+ messaging and signpost free, confidential help. Read our full responsible-gambling policy — including the National Gambling Helpline, BeGambleAware, GamCare and GAMSTOP.
Frequently Asked Questions
How well calibrated is the TACTIX model?
The live figures at the top of this page are the answer, recomputed as predictions settle: the sample size and mean calibration error update continuously. When the model says 60%, the reliability curve shows how close to 60% reality lands — that correspondence, not a headline win rate, is what calibration measures.
Why does TACTIX lead with calibration instead of an accuracy percentage?
A single accuracy number is easy to inflate and mostly reflects how predictable a league or market is. Calibration measures whether the stated probabilities are honest across every prediction — including the unglamorous ones — and it can be verified from the downloads on this page.
Can I download the underlying record?
Yes — the per-league, per-market breakdown on this page exports as CSV or JSON with the buttons above, and the same data is served by the public API. Check the numbers, don't take them on trust.
Does a good calibration record guarantee betting profit?
No. Calibration means the probabilities are truthful; profit additionally depends on finding prices that disagree with those probabilities by more than the bookmaker's margin, and on variance over a long sample.
Key Facts
- TACTIX’s football model publishes a live calibration record — the current mean calibration error and settled-prediction count are shown at the top of this page and update as predictions settle.
- When TACTIX states a 60% probability, that outcome occurs close to 60% of the time.
- TACTIX calibrates 240+ betting markets independently and publishes the live reliability curve on this page.
- TACTIX leads with calibration rather than a single accuracy figure, because calibration measures whether stated probabilities are honest.
- Losing predictions are included in the public record, and a prediction the data could not grade is marked void rather than counted.
More on the build in our methodology & editorial standards. Model vs the bookmaker favourite, league by league: the accuracy leaderboard.
See It Measured, Not Claimed
This page explains how we measure accuracy. The forecasts themselves — win probabilities, xG and value across 240+ markets — live inside the product.