Feed a language model nothing but a set of odds and ask it to find value, and you'll get confident-sounding guesses dressed up as analysis. The odds alone don't tell you why a price might be wrong — they just tell you what the market currently believes. Finding a genuine edge means grounding the analysis in the same data a professional analyst would actually look at.
SharpAI runs a two-stage process rather than a single pass. The first stage is a fast, low-cost screen across every fixture using market odds — this narrows down a large slate of matches to the handful worth a closer look. It's deliberately cheap, because most fixtures on any given matchday genuinely don't have an edge, and there's no point spending analysis budget confirming that.
The second stage is where the real work happens. Every fixture that survives the screen gets enriched with:
That enriched fixture is then re-analysed from scratch. Crucially, this second pass can overturn the first — if the deeper data doesn't actually support the initial read on a match, the pick gets dropped rather than published. That's a feature, not a failure: an odds-only screen is a filter, not a verdict.
A team that won its last three matches tells you more about its current form than a similar run from two months ago. Analysis that treats all ten recent results as equally informative will systematically lag behind a team's actual current level — especially after a managerial change, a key injury, or a run of fixture congestion. Weighting recent form more heavily isn't a cosmetic detail; it's often the difference between catching a market overreaction and missing it.
The harder part isn't generating analysis — it's knowing when not to. A match with thin data (no reliable form, no standings context) shouldn't produce a confident pick just because the model can string together plausible-sounding sentences. SharpAI's process is built to skip matches when the underlying data doesn't support a real assessment, rather than filling the gap with vague reasoning. A pick is only as good as the data behind it, and no amount of fluent writing changes that.
Every morning, the fixtures that clear both the initial screen and the deeper re-analysis get published with their reasoning attached — the standings context, the form trend, the head-to-head note, whatever specifically drove the edge. If a pick doesn't reference concrete numbers, that's worth being skeptical of.