Touchline Oracle

Methodology

An independent engine: external predictions are reference points at 5% weight and never copied. Missing data is excluded, never zero.

Model weights

Total100.0%

Weights are rescaled to total 100%. Validate any change on the Performance page using out-of-sample results.

How a forecast is built

  1. Statistical models (form, home/away, xG) each give an expected-goals pair. They are blended by weight among those with complete inputs.
  2. Context adjustments (strength, news, tactics, head-to-head, external, rest) nudge expected goals; effect scales with weight so correlated signals are not double-counted at full strength.
  3. Poisson grid over 0-10 goals per side yields 1X2, Over/Under 1.5 / 2.5 / 3.5 and BTTS.
  4. Monte Carlo runs 10,000 seeded simulations as a reproducible cross-check.
  5. Data quality rewards coverage, recorded sources, confirmed team news and fresh retrieval.
  6. Confidence blends data quality, model agreement, news reliability, simulation stability and external agreement, capped near data quality. It measures reliability, not the chance of the pick.
  7. Insufficient data is declared when no model is complete, fewer than three categories exist, no source is recorded, quality is under 40, or models conflict on thin data.
90-100 very strong80-89 strong70-79 good60-69 moderate<60 low

This tool does not browse the web. Gather figures from FBref, Understat, Sofascore, official club sites and similar, then record them with source URLs.