Known Limits

Why this page exists

We base our predictions on data we can actually measure: historical matches, odds, player ratings, head-to-head records, and recent form. But there is a lot of information we cannot observe — and pretending we can would be dishonest. This page tells you what falls outside the model so you can apply your own judgment.

What we do not model

• Player condition. Illness, injury, fatigue beyond match count, jet lag, personal issues. A player who is unwell will lose to opponents we expect them to beat.
• Equipment changes. New rubber, new blade, new shoes — these can flip a player's performance overnight and we have no visibility into them.
• Tournament context. Stakes, qualification pressure, money matches, or pre-arranged draws change the way players approach a match.
• Match-fixing risk. The leagues we cover have had documented integrity issues historically. We do not detect this directly and treat all matches as honest.
• Roster rotations. In closed-pool daily leagues (Setka, Czech Liga Pro, TT Cup) the same physical players may participate under different IDs over time. Our alias system catches some, not all.
• Officiating quirks. Umpire decisions, table conditions, ball type — we treat the physical setup as constant.
• Coaching adjustments. Tactical changes between matches are invisible to the historical record.

Things in table tennis we explicitly know are NOT real

• Home-court advantage. The "Home" and "Away" labels in the data come from the upstream feed and do not reflect any physical home advantage. Both players play on the same table under identical conditions. We tag matches with home/away purely for data linking — never as a probability boost.
• Same-day rematch lean. Our 90-day backtest on 25,000 same-day rematches showed that "back the previous winner" produces 50.2% hit rate vs 59.2% for following the market — a 9pp loss. The market already prices in the rematch effect. We surface same-day matches as context only, never as a prediction lean.
• Fatigue by match count. On 185,000 player-matches we saw no consistent drop in win rate between a player's first match of the day and their tenth. Match count alone does not predict performance drop in TT.

How we surface uncertainty

• Sample size warnings. Stat blocks render with an amber border when the underlying sample is small (e.g. win rate based on fewer than 5 matches). A 70% win rate on 4 matches reads visually different from 70% on 50.
• PASS verdict. When our model gives a probability within 5 points of 50/50, we show an explicit PASS badge instead of treating the match as actionable.
• Mixed-signals callout. When our final pick disagrees with multiple underlying signals, the match narrative names which signals dissent so you can weigh them.
• Data freshness. Each player block shows how long ago their intelligence profile was last refreshed. Stale data is flagged amber or red.

What you should not expect

We do not promise picks will hit, do not guarantee ROI, and do not tell you how much to stake. We provide one input — a calibrated probability and signal breakdown — for your own decision. Bankroll, stake sizing, and emotional control are your responsibility.