Predicify turns noisy, multi-source data into probability distributions for event contracts and sports, sizes risk with fractional Kelly, and grades its own forecasts against what actually happened. AI is used where it can be measured.
We use AI in three places, and we say plainly which are live and which are planned. In every one of them the output has to pass the same walk-forward evaluation before it can influence a decision.
Decision pipeline: where each layer touches it
Build, rule and review loop: how the system itself is made and changed
Weather is our first and deepest event-contract domain, so it is the clearest picture of how Predicify works. No single weather model is right often enough to price a contract from. We read many of them, track how each has actually performed, and combine them into a probability that a contract can be priced against. The same pattern, many independent sources weighed by track record, is how we approach every domain.
Everything settles against the official National Weather Service climate report, so we are always scored against the same answer the market uses.
Trading rules (price bands, size caps, exposure limits, when to stay out) are where AI helps most. It reads thousands of logged decisions in minutes, finds where money was lost, and drafts a rule with the evidence for and against it. It never ships one on its own.
Every decision is logged with its reason, whether we traded or passed.
AI mines the settled outcomes, splits them by city, price, lead time and phase, and finds where the losses concentrate.
AI drafts a candidate rule with the expected effect and the strongest evidence against it.
Counterfactual replay on logged history, then shadow runs, scored walk-forward.
A human approves or rejects. The specification is frozen before new data arrives.
AI implements it with tests. Every scan carries a rules-version stamp, so rule changes are never confused with model changes.
Why this matters. Most rule changes in trading systems are guesses that get remembered when they work. Here every rule has the data that motivated it, the replay that tested it, the person who approved it and a version stamp on every decision made under it. A rule that does not survive the replay never ships. Today a person approves every decision; the plan is for a person to approve the rules once, within hard limits, and for execution inside them to follow.
Every forecast follows the same path, from raw inputs to a graded decision, whether the underlying event is tomorrow's high temperature, a data release or a striker's shot count. Weather is one event-contract domain among several. The domains differ; the discipline does not.
Independent sources, timestamped at issuance, kept apart so they can be audited.
Weighted blend with explicit source-dependence handling and a spread that reflects real disagreement.
Raw model output is shrunk toward the market until history says it has earned its confidence.
Fractional Kelly with caps per position, per day and per exposure, and depth-aware fills.
Brier and log-loss, walk-forward only, with every decision and its reason logged.
Calibrated probabilities for outcomes that settle against an official or benchmark data source. Weather is our first live domain; the same pipeline is being extended to the other data-resolved categories these markets list.
Distributional models for individual player statistics, built on rolling 15-match windows. Soccer and the NFL are supported today; more leagues follow the same template.
A prediction system is only as credible as its evaluation. We build the evaluation first and treat every result as a lead until the data can carry it.
Proper scoring rules, not hit rate. A confident miss costs more than a cautious one.
Walk-forward evaluation only: a forecast never sees an outcome that was not yet known when it was issued.
Features, metrics and decision rules are written down before the data that tests them exists.
Every evaluated market is logged with the reason it was or was not traded, so questions can be answered later.
We expand one domain at a time, and each new domain has to pass the same evaluation before it is added.
Predicify is not open yet. Request early access and we will let you know as soon as it is ready.
We will only use your email to tell you when Predicify is ready.