About this app
What this is
This is a statistical prediction tool. It looks at decades of historical match data, learns patterns in how teams perform, and converts that into a probability for each possible match outcome.
It is not a crystal ball, betting advice, or financial guidance. Football is unpredictable on purpose — that's why we watch it.
How the model works
1. ELO rating
Every team has an ELO rating — a single number that summarises their strength. A team gains ELO when it wins against stronger opponents and loses ELO when it loses to weaker ones. The bigger the win, the bigger the rating change. World Cup matches change ELO the most (K=60), qualifiers a moderate amount (K=40), friendlies the least (K=20).
2. Recent form (weighted)
For each team we track the last 10 matches, with each older match weighted ~15% less than the one after it. This means yesterday's match matters far more than one from six months ago — important because squads, coaches, and tactics change. Friendly matches count for half of a competitive match in the form calculation.
3. Poisson goal model
We estimate how many goals each team is expected to score based on their attack strength, the opponent's defence strength, and the ELO gap. From these expected-goal numbers we calculate the probability of every possible scoreline using a Poisson distribution.
4. Injury overlays
We manually maintain a list of significant injuries for major national teams. When an important player is absent, we reduce that team's attack or defence multiplier accordingly — the model doesn't know who's actually on the pitch unless we tell it.
5. Knockout adjustment
Knockout matches average ~12% fewer goals than group-stage matches because teams play more defensively. When a match is flagged as a knockout fixture, the expected goals for both teams are reduced accordingly.
What "59.8% accuracy" really means
That number is the model's accuracy on a held-out portion of historical matches during training. It is not a promise of future accuracy. For comparison: a coin flip (home/draw/away) gives 33%, betting on the home team always gives ~46%, and the best public prediction systems sit around 55–62%.
Real-world performance will vary every tournament. Some weeks the model nails 4/4, others it whiffs on the headline match. That's normal — football is hard.
What the model can't do
- Predict surprises. A red card in minute 8, a goalkeeper howler, a freak deflection — none of this is in the data.
- Capture motivation. A team that's already qualified plays differently from one that needs a win to advance.
- Read live tactics. Lineups, formations, substitutions — the model doesn't see any of these.
- Account for fatigue. Travel, heat, recent fixture congestion — none of this is currently modelled.
- Predict exact scores. Even with perfect expected-goal numbers, the most likely single scoreline rarely exceeds 12% probability.
What we're working on
- Host-nation advantage as a structural feature
- Stage-context awareness (must-win vs. dead rubber)
- Better Poisson calibration for high-scoring matches
- Live track record of predictions vs. actuals (already visible in the Recent Predictions tab)
Data source
Match results are sourced from the open martj42/international_results dataset on GitHub — a community-maintained record of every international men's football match since 1872. UCL data is curated from public sources by the operator.
One more time — what this is for
FOR ENTERTAINMENT Run a fixture through the model. See what the numbers say. Argue with your friends. Watch the match. Have fun.
NOT FOR BETTING Predictions are statistical estimates. Past accuracy is not a guarantee of future results. Do not use this for any form of gambling or investment. See the Terms of Service for details.
关于本应用
这是什么
这是一个统计预测工具。它分析数十年的历史比赛数据，学习球队表现的规律，并将其转换为每种比赛结果的概率。
它不是水晶球，不是投注建议，也不是财务指导。足球的魅力本就在于其不可预测性 —— 这也是我们观看它的原因。
模型如何工作
1. ELO 评分
每支球队都有一个 ELO 评分 —— 一个总结其实力的单一数字。战胜强队会增加 ELO，输给弱队则会减少。比分差距越大，评分变化越大。世界杯比赛对 ELO 影响最大（K=60），预选赛中等（K=40），友谊赛最小（K=20）。
2. 近期状态（加权）
我们追踪每支球队的最近 10 场比赛，且越早的比赛权重越低（每往前一场降低约 15%）。这意味着昨天的比赛远比半年前的重要 —— 因为阵容、教练、战术都会变化。友谊赛在状态计算中的权重只有正式比赛的一半。
3. 泊松进球模型
我们根据进攻强度、对手防守强度和 ELO 差距，估算每支球队的预期进球数。基于这些预期进球数，我们使用泊松分布计算每种可能比分的概率。
4. 伤病叠加
我们人工维护一份主要国家队的重要伤病名单。当关键球员缺阵时，我们会相应调整该队的进攻或防守系数 —— 否则模型无法知晓真正上场的球员。
5. 淘汰赛调整
淘汰赛的平均进球数比小组赛少约 12%，因为球队踢得更保守。当某场比赛被标记为淘汰赛时，两队的预期进球数都会相应降低。
"59.8% 准确率"到底意味着什么
这个数字是模型在训练时所留出的历史比赛上的准确率，并不是对未来准确率的承诺。作为参考：随机猜测（主胜/平/客胜）准确率为 33%，永远押主队约为 46%，目前公开的最佳预测系统约为 55–62%。
实际表现因比赛而异。有时模型 4/4 全中，有时它会在关键场次失手。这是正常的 —— 足球本就难以预测。
模型做不到的事
- 预测意外。第 8 分钟的红牌、门将失误、奇怪的折射 —— 数据里都没有。
- 捕捉动机。已经出线的球队和必须取胜才能晋级的球队，踢法不同。
- 读懂临场战术。阵容、阵型、换人 —— 模型一无所知。
- 考虑疲劳。长途旅行、高温、密集赛程 —— 目前均未建模。
- 预测精确比分。即便预期进球数完美，单一比分的最大概率也很少超过 12%。
我们正在改进的方向
- 东道主优势作为独立特征
- 赛事阶段感知（必胜场 vs. 无关紧要场次）
- 对高比分比赛的泊松校准
- 预测与实际的实时对照（已在"近期预测"页面公开）
数据来源
比赛结果来自 GitHub 上的开源数据集 martj42/international_results —— 一份由社区维护的男子国际足球比赛全记录，可追溯至 1872 年。欧冠数据由运营方从公开来源整理。
再说一遍 —— 本应用是干什么的
仅供娱乐 把比赛跑一遍模型，看看数字怎么说，跟朋友辩论，看球，享受。
禁止用于赌博 预测仅为统计估计。过往准确率并不保证未来结果。请勿用于任何形式的赌博或投资。详情参见服务条款。