Prediction Analytics
Performance metrics and transparency for the 7-day price prediction engine.
Backtesting Pending
Predictions need 7 days before evaluation.
0
active predictions
0
regional + grid predictions
Current National Predictions
Live 7-day price direction forecasts for each fuel type
No Active Predictions
Predictions are computed every 30 minutes. Check back soon.
Accuracy Over Time
Daily direction accuracy from backtested predictions
Loading data...
Failed to load data
No backtesting data yet
Predictions need 7 days before they can be evaluated
Predicted vs Actual
Each dot is one evaluated prediction (closer to diagonal = more accurate)
Loading data...
Failed to load data
No evaluated predictions yet
Scatter data will appear once predictions have been backtested
Accuracy by Confidence
Do higher confidence predictions actually perform better?
Loading data...
Failed to load data
No confidence data yet
This chart requires evaluated predictions across confidence levels
Action Outcomes
How each recommendation type has performed
Loading data...
Failed to load data
No action outcome data yet
Results will appear once predictions have been evaluated against actual prices
Action Distribution
Current prediction recommendations
No active predictions
Confidence Distribution
Signal strength across all predictions
Signal Weights
How each signal contributes to predictions
A robust regression over recent daily averages: which way prices are moving and how fast.
Weekly repricing rhythm. Live scoring caught this signal misfiring in mid-2026, so it carries a token weight while it is reworked.
Supermarkets reprice ahead of the major brands; a widening gap between the two often comes before a broader move.
Whether a local area has caught up with the national average yet. Local prices tend to follow it within a week or two.
Price changes travel geographically. This watches neighbouring areas for a move heading your way.
How long prices have held still. A long-frozen price is more likely to move soon, so this scales the engine's conviction.
Brent crude oil, in sterling. The only signal watching the outside world: pump prices follow the oil market with a lag of roughly four to eight days, measured against the site's own price history.
How the Predictions Work
Multi-Signal Analysis
7 signals (trend, seasonality, brand behaviour, national & regional momentum, price stickiness, and Brent crude oil) are computed and weighted. Several share the same underlying price history.
7-Day Horizon
Each prediction forecasts the direction and magnitude of price changes over the next 7 days for a specific fuel type and region.
Continuous Backtesting
A daily forecast is archived for each fuel and area. After 7 days, predicted and observed outcomes are compared using the same stations.
Graceful Degradation
Signals that lack sufficient data are automatically disabled. The confidence score describes signal quality and agreement; it is not a measured probability of being right.
Regional Granularity
Predictions run at national, postcode area, and geographic grid levels, capturing local price dynamics that differ from national trends.
Actionable Advice
Buying suggestions are withheld when the evidence is weak. Suggested windows are guidance from the seven-day forecast, rather than a measured best time to buy.
Engine Updates
How the prediction methodology has changed, and why
July 2026
Crude oil signal and a full reweighting
Brent crude oil joined the engine as its first external signal, after a study of the site's own price history showed pump prices follow the oil market with a lag of roughly four to eight days. Every archived prediction was then scored against what prices actually did, all signal weights were rebalanced against 2.5 million of those evaluations, and the day-of-week signal was cut to a token weight after the scoring showed it misfiring. Replay testing used the same months as the reweighting, so its results are development evidence and do not establish future accuracy.
Ongoing
These figures use daily national forecasts from the current engine over the most recent 90 days. Earlier engine versions are excluded. The same stations are compared at the forecast time and 7 days later; incomplete station coverage is left unscored. Results appear as forecasts mature. Overlapping seven-day forecasts are not independent trials, and direction accuracy does not measure savings from a buying suggestion.
Prediction Coverage
Predictions are computed across multiple geographic scales