"85% accurate!" is the easiest claim in football predictions to make, and the least useful one to read. Accurate at what? Over which matches? Since when? Goal Guru's new Track record is our answer. Instead of one number on a website, the app grades every prediction after its match and shows the results in several ways, so you can judge them yourself. Here's how it works, and why it's built that way.
Why One Accuracy Number Tells You Almost Nothing
Some predictions are easy. Most matches have at least one goal, so a model that predicts "over 0.5 goals" every time will be right very often without knowing anything. Predicting the result of a match between a title contender and a relegation side is easier than calling a mid-table derby. A single accuracy figure blends all of this together, and it can be pushed up simply by choosing which markets and matches to count.
That's why the Track record never gives you one figure on its own. It breaks the record down in the ways that make it hard to flatter.
Every Prediction, Graded
Every pre-match prediction the model makes is graded once its match has finished. Grading runs once a day, the morning after the matches, so the record always covers finished games with settled results. Every prediction counts, whether it was right or wrong.
You choose the period: a single day, the last 7, 30, or 90 days, or all time. One day shows how a matchday went, and can swing a lot. Longer periods smooth out a few lucky or unlucky results and show the model's real level.
By Market
Accuracy is shown separately for each prediction type: match result, both teams to score, each goal line, and so on. Each market answers its own question, and they shouldn't be averaged into one.
A market's rate only appears once it has at least 30 graded matches. With fewer, two or three results either way would swing the percentage wildly. Showing that as if it meant something would be the kind of flattering number we're trying to avoid.
By Confidence: Sure Picks and Strongest Picks
When the model is very confident, how often is it right? Sure Picks answers that directly: the predictions the model gave at least a set probability, and how many of them won. Strongest Picks lists the model's most confident calls one by one, next to what actually happened. You can see the individual hits and misses, not just a rate.
The View That Matters Most: Did 70% Mean 70%?
Here's the subtle part. A probability isn't a yes-or-no call. If the model says an outcome has a 70% chance, it's also saying it expects to be "wrong" about 30% of the time. Judging that 70% only by whether it came true misses the point. The real question is whether, across many matches where the model said 70%, the outcome happened about 70% of the time.
Statisticians call this calibration, and it's what the How often each chance came true view shows. Tap a market and the graded picks are grouped by the chance the model gave them. For each group:
- The tick marks the chance the model gave.
- The bar shows how often those outcomes actually happened.
When the bars sit close to their ticks, the model's numbers mean what they say, and you can use them as real probabilities. If bars sat consistently below their ticks, the model would be overconfident. If they sat above, it would be too cautious. This view is available for a single day and for 7, 30, and 90 days.
Calibration is also why a well-calibrated model can look "less accurate" than a flashy claim. It's honest about uncertainty. A 55% call that comes in 55% of the time is doing its job, even though it's "wrong" nearly half the time.
What's Not in the Record Yet
Value picks are outcomes the model rates as likely, at least 50%, and more likely than the bookmaker's price implies. We recently tightened that definition. Results under the new rule are graded every day, but they won't appear in the Track record until there are enough of them to judge fairly. Publishing a few days of results would be exactly the kind of small-sample number this page argues against.
How to Use It
- Before trusting a market, check its own record, not the overall figure.
- Prefer the 30- and 90-day views over a single day when judging the model.
- Use the chance view to read predictions properly. If 70% calls have been coming in about 70% of the time, treat the next 70% as roughly a 7-in-10 chance, not a near-certainty.
- Combine it with your own conditions. A pre-match alert can require both a model probability and a form or head-to-head condition, so the model is one input rather than the only one.
Find the Track record in Guru Vision, next to the Today, Live, and Surprises tabs. It's the same principle as backtesting your Smart Alerts: trust what you can check, not what you're told.