SoloPredict, and how to audit a published record

The five checks

These apply to any prediction site claiming a record, ours included. None of them require statistical training, and together they catch almost every way a record can be quietly flattering.

1. Was the pick timestamped before kick-off? A record assembled after results are known is not a record, it is a list. If there is no visible publication time earlier than the match, nothing else on the page can be trusted.
2. Are the odds recorded next to each pick? Without the price, a hit rate cannot be converted into whether following it made or lost money, which is the only question that matters. Records that omit odds are omitting the half that could be unflattering.
3. Do the losses appear at the right rate? If a site claims 65% and you scroll a hundred entries, you should see roughly thirty-five losses. If losses are conspicuously rarer than the headline implies, the headline is counting something other than what is displayed.
4. Are voids marked rather than removed? Postponements and abandonments happen constantly. A record that silently drops them has a quiet mechanism for dropping other things too. Marking them void and excluding them from the percentage is the honest handling, and it is what we do.
5. Do the totals reconcile? Count how many picks were published in a period and how many appear in the settled record for that period. The two numbers should match. When they do not, the gap is the story.

Our own record at /accuracy is built to pass these, and it is worth saying plainly that it is currently short, because the site is new. A short record is a weak claim. It is still a better claim than a long one you cannot audit.

What the machine-learning badge does and does not tell you

SoloPredict markets an AI product alongside its standard tips, and the term deserves a straight explanation rather than either hype or dismissal.

Machine learning is a legitimate approach to football forecasting. You take a large history of fixtures with their inputs and their outcomes, and you fit a model that learns which patterns of inputs preceded which results. Done properly it can find relationships a hand-built formula would miss. It is not magic and it is not nonsense.

The problem is that the label carries no information about quality. The questions that determine whether an ML forecast is any good are all invisible from the outside:

• Was it tested on matches it had never seen? A model evaluated on the same data it learned from will look superb and predict nothing. This is the single most common failure and it produces spectacular reported accuracy.
• Was the test set later in time than the training set? Football moves. Testing on randomly shuffled historical matches lets the model learn from the future, which no live forecast can do.
• How many inputs, against how many matches? Enough variables will fit any history perfectly and generalise to nothing.
• Is it calibrated? When it says 70%, does that happen about 70% of the time? A model can rank matches well and still be badly wrong about the probabilities it attaches.

None of those answers are published by anyone in this market, which means “AI predictions” should move your confidence exactly as much as “statistical predictions” does, which is somewhat, and only alongside a record you can check. The same reasoning is why we say crash-game predictors cannot work, although that case is stronger still: there the outcome is generated from a secret the software never sees, so no amount of modelling touches it.

The paid tier, and why we would not buy it

SoloPredict sells predictions as well as giving them away. Our position on paid tips generally is that they are a bad purchase, and the reasoning is arithmetic rather than suspicion.

A bookmaker prices with a margin built in. To profit from tips you must first beat that margin, and then also earn back what you paid for the tips on top. Suppose a service costs enough to represent 5% of your monthly staking. Your selections now need to overcome the bookmaker's edge and clear a further five points before you are level. Edges of that size are rare enough that people who have them do not advertise.

Which leads to the question worth sitting with: if a service genuinely held a reliable edge, why would it retail that edge rather than stake it? Staking scales without customers, support, refunds or marketing. The honest answers are that the edge is small and uncertain, or that selling subscriptions is simply the more dependable revenue. Neither is a reason to buy. The harder-edged version of this argument, applied to people selling certainty rather than probability, is on our fixed matches page.

What we would actually use it for

The free predictions are worth reading the way any competent model is worth reading: as a second opinion that has looked at form, goals and availability without any emotional attachment to a team. That is a real service and it costs nothing.

Read the probability rather than the selection where one is given. Compare it against the price you can actually get, using implied probability, and ignore anything where the model and the market agree, because there is no disagreement to profit from. And treat the record as the thing that makes the rest meaningful, which is why it is worth the ten minutes to audit it properly rather than trusting the summary at the top.