Methodology

Seer9 starts with quantitative probability models. It may then compare those estimates with executable prediction-market prices and source-attributed research signals. The system is designed to preserve the numerical model as the anchor rather than allowing a language model to invent a probability from prose.

Seer9 does not predict the future with certainty. It estimates probabilities and measures how those estimates perform after outcomes resolve.

Each supported sport has its own numerical pipeline. MLB forecasts use game-state and matchup inputs through a baseball probability engine. Tennis forecasts use player, surface, serve, return, and live score state through a point-to-match Markov model. Inputs are frozen with each forecast so later evaluation can distinguish model error from changed information. The model version is recorded with each public forecast.

The product does not use one generic language-model opinion for every sport. Sport-specific probability logic remains the anchor.

• Soccer: team ratings, expected goals, home advantage and available pregame context produce match-winner probabilities, including the draw. Sparse team evidence raises uncertainty.
• MLB: starting pitchers, team rates and matchup inputs feed a baseball probability engine. A late starter change can invalidate the original context.
• WNBA: pregame team records are shrunk toward an even matchup before a home adjustment. Live estimates also use score, time remaining and recent play data.
• Tennis: player, surface, serve and return estimates feed a point-to-match model. Live estimates can incorporate the observed score.
• NFL: team ratings feed a drive-level game simulation. The published pregame call records the simulation version and expected score.
• UFC / MMA: dated, opponent-adjusted fight results and a restrained pre-UFC record signal estimate the matchup. Limited histories and inactivity widen uncertainty.

Where matching contracts are available, Seer9 compares its model probability with public Kalshi or Polymarket prices. A displayed gap is the difference between two estimates, not a promise of profit.

Public prices may be delayed, thinly traded, or not executable at the displayed level. Users should independently verify a market before making any decision.

A low-cost DeepSeek model may review structured, source-attributed evidence for information the quantitative model has not encoded, such as a confirmed absence or material late-breaking context. The research layer receives the numerical anchor and operates under adjustment limits.

Unsupported claims, missing information, and scenario risks are recorded. Language-model output is not treated as ground truth, and model changes are not promoted solely because an explanation sounds persuasive.

The public Predictions board shows qualifying pregame calls in MLB, WNBA, tennis, NFL and UFC / MMA, with up to three qualifying free calls per game day. Soccer releases one free match forecast per active league slate and publishes graded historical selections, accuracy, calibration and paper strategies after settlement. The complete slate and Parlay Board are subscriber features. Parlays combine uncertain legs, so correlation and compounding error can lower their hit rate.

A public call keeps its publication time, evidence cutoff, probability, original factors and caveats. The result is added only after an official settlement. Wins and losses both remain visible in the match record; an unsettled event is never a win.

War Room updates the decision context for supported live events. The live numerical engine resumes from observed game or match state when data is available. Freshness indicators matter: delayed source data can produce a stale view even when the page is open.

Macro Flow analyzes public prediction-market transactions and wallet behavior. It may aggregate related fills to show a trader's broader position rather than mistaking each fill for a separate conviction.

Public wallet activity does not reveal every hedge, motive, or off-platform position. A large trade is a research signal, not a recommendation or proof that a wallet is informed.

Strategy Lab freezes eligible forecasts, groups parlay legs, and grades them after official outcomes become available. Macro Flow paper positions record simulated entries, exits, and settlement without placing real orders.

Settled misses remain visible internally for evaluation. Records are not silently rewritten after an outcome, and unsettled forecasts are not counted as wins.

Seer9 evaluates probability quality with outcome-aware measures such as Brier score, log loss, calibration, closing-price movement where available, and hit/miss records. Evaluation is separated by sport, market, horizon, and confidence so one strong segment cannot conceal a weak one.

Historical results can suggest a candidate adjustment, but production behavior should change only after enough representative data, out-of-sample checks, and rollback criteria. The learning loop supports model governance; it does not guarantee continuous improvement.

• Third-party feeds can be late, incomplete, or unavailable.
• Market prices can move before Seer9 refreshes.
• Models cannot encode every injury, matchup, incentive, or event.
• Language models can misunderstand evidence or produce false claims.
• Backtests can overstate performance when conditions change.
• Paper execution does not reproduce every real spread, fee, or fill.

• Accept wagers or customer deposits
• Custody money, securities, contracts, or exchange credentials
• Execute or route trades
• Guarantee an outcome, profit, or level of accuracy
• Provide financial, investment, legal, tax, or gambling advice
• Take a percentage of user winnings