Verified Tennis Trading Results

How We Generate Tennis Trading Signals

Every signal published on fault.bet originates from a quantitative model that blends gradient-boosted classification with point-level serve simulation. The system was built to answer a single question: is the Betfair back price for this player higher than it should be? When the answer is yes, and the margin is large enough, a signal is generated. Below we explain the full methodology so you can evaluate the track record on this page with complete context.

Model Architecture: LightGBM + Markov Chain Blend

The core prediction engine uses LightGBM, a gradient-boosted decision tree framework, to estimate the probability that each player wins a given match. LightGBM was chosen for its speed, its native handling of categorical features like surface type and tournament tier, and its resistance to overfitting when properly regularised. The model is trained on 1.6 million historical match records spanning ATP, WTA, Challenger and ITF circuits from 2005 to the present day.

Running alongside the LightGBM classifier is a Markov chain serve model that simulates matches at the point level. This component ingests 562,000 point-level serve records and constructs transition matrices for each player on each surface. It calculates the probability of holding serve, breaking serve and winning tie-breaks, then runs thousands of Monte Carlo simulations to arrive at a match-win probability.

The final probability is a weighted blend of both outputs. The LightGBM prediction captures broad patterns such as surface form, recent momentum, fatigue and head-to-head tendencies. The Markov chain output adds granularity by modelling serve mechanics, return effectiveness and clutch performance under pressure. By combining both, the model achieves a more robust probability estimate than either component alone.

Feature Engineering and Data Sources

The LightGBM model uses 25 engineered features per match. These include rolling Elo ratings segmented by surface, weighted head-to-head records with recency decay, tournament-level fatigue indices derived from scheduling data, travel distances between consecutive events, and aggregated serve statistics across multiple return windows (last 10, 20 and 50 matches). Additional features capture weather-correlated performance on outdoor hard courts, altitude adjustments, and a categorical encoding of the tournament round.

All data is sourced from publicly available ATP and WTA match databases, supplemented with point-level serve records from the Match Charting Project and proprietary scraping of tournament draw sheets. Historical Betfair odds are recorded via the Betfair API at the moment each signal is generated, ensuring that every price on this results page was genuinely available in a live market.

Signal Criteria: When a Signal Fires

Not every match where the model disagrees with the market becomes a signal. Three conditions must all be met before a signal is published:

1. Betfair market confirmed — The match must have an active Betfair market with sufficient liquidity. Matches without a Betfair market or with less than £5,000 matched at signal time are excluded.
2. Edge of 6% or greater — The model probability must exceed the implied Betfair probability by at least 6 percentage points. For example, if the Betfair price implies a 50% chance, the model must rate the player at 56% or higher. This threshold filters out marginal edges that would not survive commission and price drift.
3. Confidence score of 40 or above — The confidence score is a composite metric reflecting data availability, feature completeness and model agreement between the LightGBM and Markov chain components. A confidence of 40 indicates moderate certainty. Scores above 80 indicate strong agreement between both model branches with deep historical data available.

When all three conditions are met, the signal is timestamped, the Betfair price is recorded, and it is published to subscribers via Telegram and the fault.bet dashboard. From that moment, the signal is locked in and tracked on this page regardless of outcome.

faultbet Staking

Each signal receives a stake proportional to the detected edge, adjusted for signal confidence and data quality. Higher conviction signals with cleaner underlying data receive larger allocations. Lower confidence signals are automatically scaled down to minimum stakes.

Stakes range from 0.15% to 2.5% of bankroll. The exact allocation factors in the edge size, the market price, our confidence in the signal, and the quality of the underlying player data. This ensures we size up on our strongest calls and protect capital on weaker ones.

Paper Trading

All results shown are from paper trading. Prices are recorded from live Betfair markets at signal time and settled against official match results. Results do not account for Betfair commission (2-5%), price drift, or liquidity constraints. Past performance does not guarantee future results. 18+ BeGambleAware.org.