Understanding Signal Accuracy
Learn about our signal methodology, the challenges we face, and how we honestly track every Confirm and Contradict read across our main markets.
The Challenge: Scale and Complexity
Table tennis is one of the most active sports in terms of match volume. On any given day, we analyze matches across dozens of leagues, with hundreds of players competing in multiple tournaments simultaneously.
Multiple Leagues
We track matches from Czech Liga Pro, Setka Cup, TT Cup, TT Elite Series, Challenger Series, and many more.
High Match Volume
Players often compete in multiple matches per day, sometimes across different leagues and tournaments.
Three Main Markets
We analyze Match Winner (1x2), First Set, and Over/Under — each with dedicated signal calibration.
Our Signal Methodology
Multi-Signal Consensus System
Our system runs independent signals on every match, built on 470K+ historical matches with clean historical data:
- Signal Consensus: Independent signals (Glicko-2, Form L5/L10, Streak, H2H, Market, Last Activity, Odds Role, First Set) vote on each match
- League Portrait: Intra-day tracking of which signals are working right now per league — stops signals that aren't contributing
- Clean Historical Data: Signals are computed strictly from data available at match time — no reprocessing or future data leakage
First Set Calibration
First Set predictions use H2H-first statistical blending calibrated by sample size:
- H2H-first blend: 80% H2H / 20% empirical when 12+ H2H samples available
- Low-spread shrinkage toward 50% for close matchups to reduce noise
- Requires minimum 3 H2H matches to generate reliable First Set probabilities
Over/Under Statistical Model
For Over/Under predictions, we use a statistical mixture model that combines signals from H2H history and match context:
- Multiple Signals: H2H scoring history, average points, MW spread, player tendency, scoreline distribution, and line-level correction
- Probability Calibration: Predictions are shrunk toward neutral to reduce overconfidence in small-sample matchups
- Over/Under Value Bets are flagged when our signal probability differs meaningfully from the implied odds — a signal read, not a guarantee.
Three Main Prediction Markets
The Underdog Factor: Why Better Players Sometimes Lose
Understanding Upsets
One of the most challenging aspects of table tennis predictions is that underdogs frequently win, even when they appear to be the weaker player on paper. This happens for several reasons:
1. Form Over Reputation
A lower-ranked player who is in excellent form can defeat a higher-ranked player who is struggling. Our Form L5 and Form L10 signals track recent performance more heavily than overall ranking.
2. Match Context Matters
Tournament importance, fatigue from previous matches, and motivation levels can significantly impact outcomes. A player fighting to avoid relegation may perform better than a player who has already secured their position.
3. Playing Style Matchups
Some players have styles that are particularly effective against certain opponents, regardless of overall ranking. Our H2H analysis helps identify these patterns, but they're not always obvious from surface-level statistics.
4. Variance in Short Formats
Table tennis matches can be decided by small margins. A few key points can swing an entire match, especially in best-of-5 or best-of-7 formats. This inherent variance means upsets are more common than in sports with longer formats.
5. Incomplete Data
While our signals cover key dimensions, some factors are difficult to quantify: player injuries, personal circumstances, equipment changes, or tactical adjustments. These can significantly impact performance but aren't always visible in the data.
Current Accuracy Numbers
These numbers reflect our live prediction performance — measured on real matches, not backtested data:
Overall accuracy across all matches and markets
Confirm read track record — our highest-conviction calls where signals agree with the market favorite
Important Context:
- • Edge varies by market type (Match Winner, First Set, Over/Under)
- • Confirm and Contradict tiers each have their own published record — check the Performance page for current numbers
- • Value Bets are detected when our signal consensus has a meaningful edge over implied odds
- • Some leagues (e.g. TT Cup) are structurally harder to predict due to format
- • We're transparent about our limitations and always show confidence levels
Data Quality and Limitations
For each match, we collect and analyze comprehensive data:
- • Head-to-head records (last5, last10, last20, last50, last100, last6Months, all-time)
- • Temporal win rates across the same windows
- • Performance by league and tournament type
- • Bookmaker odds (used as a feature in the core model)
- • Set-by-set win rates (S1–S5) from H2H history
- • Total points average, min, max, median from H2H
- • Scoreline distribution (3-0, 3-1, 3-2 percentages)
Despite our comprehensive analysis, some factors are beyond our control:
- • Last-minute injuries or withdrawals
- • Equipment changes or technical issues
- • Personal circumstances affecting motivation
- • Tactical surprises or unexpected strategy changes
- • Environmental factors (venue, crowd, conditions)
- • Human error or referee decisions
How to Use Our Predictions
Consider Confidence Levels
Pay attention to the Confirm and Contradict tier indicators and each one's public track record before following.
Review Multiple Markets
Don't rely solely on Match Winner. Check Handicap, Over/Under, and First Game predictions for a more complete picture.
Understand the Context
Review H2H records, recent form, and match importance. Our predictions are data-driven, but context matters.
Account for Variance
Upsets happen in table tennis more often than in other sports. Even our highest-conviction Confirm reads will not win every bet. Use proper bankroll management.
Combine with Value Bet Analysis
Use our Value Bets filter to find matches where our signal consensus suggests the odds undervalue the true probability.
Continuous Improvement
We're constantly working to improve our edge detection:
- • League Portrait refreshes intra-day as signal performance evolves
- • Refining signal thresholds based on per-league live, point-in-time results
- • Improving data collection and historical data quality
- • Validating new signal patterns before surfacing them as tiers
- • Learning from edge detection misses to identify new patterns
Point-in-Time Records — No Hindsight
Every prediction is snapshotted the moment it is published and graded only after the match officially settles. Cancelled and postponed matches are voided, not counted. Nothing is edited retroactively — the record you see is the record as it happened.
The Market Is the Benchmark
Raw hit rate is a vanity metric: picking every heavy favorite produces a high win percentage and still loses money to the vig. The honest question is whether calls outperform the probability already implied by the odds. That is the comparison we track and publish, league by league.
Variance and Sample Size
Seven-day windows swing wildly — a run of coin-flip matches can make any system look brilliant or broken. Judge performance on 30- and 90-day windows with real sample sizes. One hot week means nothing; one cold week means just as little.
Where to See the Ledger
The Track Record page shows daily results, per-league and per-tier breakdowns, and both sides of the record — the wins and the losses. That page, not a highlight reel, is how we ask to be judged.