How we make projections: and how we grade them in public.
Nebula Insights publishes an independent model number beside the market's, logs it before the event, and grades it after against the same games the market priced. This page explains exactly how that works and shows the honest current numbers, including where we trail. Information, not picks.
Methodology updated: July 2026
How the projections are made
Our models learn from years of historical games. They use only team, player, park and schedule data team and pitcher form, park factors, rest and travel, and similar public inputs. Crucially, they never see a betting line: the projection is a genuinely independent estimate, not a repackaging of the market.
Team & pitcher form, park factors, rest and schedule, built as of game time so nothing from the future leaks in.
Walk-forward across seasons (2021 to today), so every test game is one the model had not seen.
Outputs are calibrated so a stated 55% actually wins ~55% of the time, no false confidence. Win probabilities are isotonic-calibrated and clipped to a 2%-98% range; we never publish a game as a certainty.
We show our number next to the book's, never a lean, arrow, or bet direction.
For MLB we publish a projected total (the over/under number) and a calibrated home-win probability for the moneyline. For the run line and for NFL and UFC, we publish the market's own numbers and factual context only, because our testing found no reliable signal to add there, and we would rather show nothing than dress up noise.
The honest numbers
Measured out-of-sample from 2021 to mid-2026, on the same games that had a real opening line. We show our error next to the market's so you can judge the model yourself.
An improvement story, honestly told. Our previous totals model missed by 3.72 runs on average; the current one misses by 3.52. That is real progress against our old self, and it still trails the market. Both facts are true, and we say both.
These are the only model performance numbers we quote publicly. Every figure above is fixed to our audited July 2026 evaluation; we do not round them favorably or restate them.
How the track record grades itself
The public track record is the trust asset. It works on four rules, and nothing about it is discretionary:
- Logged before the event. Every projection is written down and time-stamped before first pitch. You are seeing what we committed to, not a story told after the fact.
- Graded after the result. Once the game is final we fetch the score and record the miss, automatically, on a schedule.
- Identical, genuinely-lined games. We compare our error to the market's error on the exact same games that had a real opening line. No cherry-picking the games that flatter us. A block of early totals rows logged between 2026-07-16 and 2026-07-29 has no recorded opening line at all; those games are excluded from every line-relative comparison, and the excluded count is published beside the statistics rather than quietly dropped.
- Immutable snapshots. A logged projection is never backfilled or edited. If we were wrong, the record stays wrong. What we published is what you see.
For markets where we publish no projection (the run line, NFL, UFC), the track record shows the market's own benchmark instead, clearly labelled as the sportsbooks' record, not ours.
What this is: and is not
Nebula Insights is a free, independent analytics layer. We show the math and the receipts. We are never paid when you lose, and we run no sportsbook affiliate links.
Frequently asked
How are the projections made?
How are projections graded?
Why do you show your losses?
Is your projection better than the market or Vegas?
Do you sell picks?
Nebula Insights content is provided for informational and analytical purposes only. It is not betting advice, not financial advice, and not a guarantee of any outcome. Odds move; verify all prices and availability directly with sportsbooks. 18+ only. If gambling is a problem for you, call 1-800-GAMBLER.