How Bozzle Bets works.
Bozzle is a player-prop research system. It scans available markets, builds sport-specific projections, applies context and quality gates, keeps permanent prediction receipts, and helps narrow a large slate into a smaller set of decisions worth reviewing.
Quick start
The normal workflow is intentionally simple:
Core concepts
Raw model probability
The probability produced by the sport/prop-family statistical model before Bozzle applies uncertainty and calibration shrinkage.
Bozzle probability
The final displayed probability after reliability, sample quality, market breadth, player status and other safeguards are considered.
Threshold
The minimum probability used to qualify a signal for tracking. The default is 65%. Crossing the threshold does not automatically mean PICK.
Context completeness
A 0–100 score describing how much of the expected supporting information is available. It is a data-quality / evidence score, not another probability.
Market consensus
A median or reference constructed from sportsbook prop lines returned by the data provider. More independent books generally make the reference more useful.
Correlation cluster
A group of predictions that can win or lose together because they depend on the same player/game mechanism. Clusters prevent related props from pretending to be independent evidence.
Decision labels
Daily Scan
Daily Scan is the broad discovery layer. Instead of researching one player at a time, Bozzle loads the scheduled games, supported player markets, available roster/lineup information and required player history, then scores the board automatically.
What Daily Scan does
- Finds the games on the selected date.
- Loads the supported sportsbook player-prop board.
- Resolves players to the correct game/team.
- Loads the historical sample required by that sport's engine.
- Scores supported lines and ranks the strongest results.
- Freezes qualifying live pregame signals into the Model Ledger.
- Leaves unsupported, suspicious or incomplete markets visible without force-scoring them.
Why the number of “scored props” can be much larger than the shortlist
Scanning is intentionally broad. Qualification, PICK gates, Focus, correlation controls and Bankroll Builder are progressively narrower layers. A large scan followed by a small shortlist is expected behavior.
Exact Line / Player Hub
Daily Scan normally evaluates sportsbook consensus. Fantasy-pick platforms can display a different number. Exact Line exists so the final analysis is performed against the number you can actually select.
- Choose the game and player, or open a candidate directly from Daily Scan.
- Select the prop.
- Replace the sportsbook reference with the exact line shown on the third-party platform.
- Run Analyze Exact Line.
- Read the calibrated probability, decision, projection, market comparison, context receipt and warnings.
A seemingly small line change can materially change the probability. For example, a sportsbook line of 251.5 and a fantasy-platform line of 257.5 are different statistical questions and should not share the same final probability.
Opening-only markets
If a market is labeled OPENING ONLY, Bozzle has an opener but not a trustworthy current live reference. It will not promote that as if it were a current consensus market.
Prediction engines
Bozzle is designed as one platform with separate sport-specific engines. NFL logic is not reused for MLB, and future NBA logic is intended to have its own model family.
NFL engine
The NFL model combines leak-free historical performance, recent form, opportunity/efficiency components where available, sportsbook market information and structured context. Examples include carries and yards per carry for rushing, targets/catch rate/yards per reception for receiving, and attempts/efficiency for passing.
- Historical games are filtered so a target game cannot leak its own final stats into a pregame prediction.
- Current sportsbook market breadth and opening/current lines are used as evidence, not treated as truth.
- Availability/designation risk can force REVIEW or AVOID.
- Context can include workload, team/opponent profiles, game betting environment, expected script and NFL venue/weather information.
- Outdoor NFL weather is sourced through U.S. National Weather Service data when the venue/forecast can be resolved.
MLB engine
MLB uses separate hitter and pitcher families and baseball-native count/event distributions. The current model does not treat every prop like a continuous normal variable.
Binomial/event family
Used for event-per-opportunity markets such as hitter hits, walks and batter strikeouts when the event structure is appropriate.
Poisson / negative-binomial family
Used for count outcomes such as pitcher strikeouts, pitcher hits/walks allowed, earned runs and selected hitter count stats. Negative binomial is preferred when observed variance exceeds a simple Poisson assumption.
Shifted empirical family
Used for lumpy distributions such as total bases, combination hitter stats and pitcher outs, where recent observed outcome shapes are useful.
The MLB baseline incorporates opportunity (plate appearances or batters faced), recent production rates, historical form and sportsbook information. Exact Line can also retrieve deeper context such as handedness splits, batter-versus-probable-starter samples and pitch-type profiles when available.
Calibration
Bozzle deliberately shrinks raw model confidence toward 50% when evidence is thinner. Larger samples, confirmed lineups/starts and broader markets can retain more of the raw edge. Model Lab is responsible for determining whether those displayed confidence levels are actually calibrated over time.
Bozzle Focus
Focus is the MLB short-listing layer. The complete threshold board remains available, but the default Focus view is intentionally selective.
A Focus candidate must first be a true PICK and then clear stronger requirements such as probability, mature sample size, confirmed lineup/probable starter, context quality, sportsbook breadth, market agreement, low player-status risk and evidence reliability.
Focus also tries to avoid showing multiple highly related props from the same player/correlation cluster. The goal is not to create more picks; it is to reduce a large research board into a manageable list worth deeper review.
My Prospects
My Prospects is a private scratchpad. Use + Prospect when a signal deserves another look but is not necessarily an official target or recommended entry.
- Stores the exact player, stat, line, side, probability and supporting snapshot.
- Allows finalists to be marked and reviewed later.
- Can warn about low context, thin markets or related exposure.
- Is separate from the Model Ledger and separate from Bozzfather.
Current Prospects are browser-local. Clearing browser/site storage can remove them.
Model Ledger: Bozzle keeps receipts
The Model Ledger exists to prevent cherry-picking. A prediction should not count only because somebody liked it after the fact.
What gets frozen
Qualifying live pregame signals that cross the tracking threshold are frozen before the event starts, even if the user never adds them to Prospects or publishes them as a Bozzfather target.
What a receipt contains
Typical fields include sport, player, game, stat, line, side, model probability, projection, model version, market context, context completeness, timestamp and settlement state.
Model versions are separate predictions
The model version is part of prediction identity. A v0.2 and v0.3 prediction for the same player/line can coexist so Model Lab can compare model generations fairly.
Settlement
After a game becomes final, Bozzle can retrieve the player's final stat and grade the frozen receipt as WIN, LOSS or PUSH. If the stat cannot be resolved reliably, the record is held for verification rather than guessed.
Model Lab
Model Lab is the research dashboard for settled Model Ledger receipts. It is designed to answer whether the model's confidence is justified, not merely whether a handful of picks won.
Hit rate
Wins divided by settled wins + losses. Pushes are not treated as wins or losses.
Average confidence
The average displayed model probability across the settled sample.
Calibration gap
Actual hit rate minus average predicted probability. A large persistent gap may indicate over- or under-confidence.
Brier score
A proper probability scoring rule: average squared error between the forecast probability and the binary outcome. Lower is better.
Sample strength
Current labels are EARLY under 30 settled, DEVELOPING at 30+, USEFUL at 75+, and STRONG at 150+.
Independent clusters
A correlation-adjusted view that retains receipts but uses one representative signal per correlated player/game cluster when estimating independent evidence.
Model Lab also breaks performance down by prop type, context bands and model version. It can surface research recommendations, but it does not silently rewrite model weights.
Bozzfather
Bozzfather is the public official-target layer. It is intentionally separate from Daily Scan and Prospects.
- Only Publisher mode can publish official targets.
- The public page displays at most the strongest three active targets for the day.
- When nothing has been published, The chair stays quiet.
- Published targets carry probability, line and supporting evidence where available.
- Official history shows settled wins/losses and the public target record.
Bozzfather is not supposed to publish everything the model tracks. The Model Ledger is exhaustive accountability; Bozzfather is selective publication.
Bankroll Builder
Bankroll Builder is a private decision/risk layer. It does not create new projections; it consumes today's current calibrated Bozzle signals.
Step 1 — Enter your settings
- Bankroll: the dedicated amount you have chosen for this activity, not general living money.
- Base unit: converts the suggested dollar size into units. A 1% unit on a $500 bankroll is $5.
- Risk sizing: Conservative uses ¼ Kelly; Standard uses ⅓ Kelly.
- Multiplier: enter the actual total-return multiplier displayed by the platform for the entry type you are considering.
Step 2 — Eligibility gate
Builder starts with 65%+ live signals, then excludes legs that fail its stricter combination rules. Reasons can include REVIEW/PASS status, availability risk, low context, fewer than two reporting books, material sportsbook disagreement, unconfirmed MLB lineup/start or low evidence reliability.
Step 3 — Correlation controls
Builder hard-blocks same-player and same-correlation-cluster combinations. It also blocks same-game combinations until Bozzle has a validated covariance/correlation model for those relationships.
Step 4 — EV calculation
Break-even probability = 1 / total-return multiplierEstimated EV per $1 = adjusted joint probability × multiplier − 1For example, a 3.0× all-or-nothing total return has a break-even probability of 33.3%. Bozzle multiplies the calibrated leg probabilities only after known correlation blocks, then applies an additional model-risk haircut before computing EV.
Step 5 — Fractional Kelly sizing
For an all-or-nothing return multiplier M, the Builder first computes full Kelly from the adjusted joint win probability, then uses only a fraction of it:
b = M − 1Full Kelly = (b × p − (1 − p)) / bSuggested fraction = Full Kelly × ¼ (Conservative) or × ⅓ (Standard)Additional caps are then applied. Current per-entry caps are deliberately tighter than the Kelly result:
| Mode | 2-pick cap | 3-pick cap | Total suggested slate cap |
|---|---|---|---|
| Conservative | 1.0% of bankroll | 0.75% | 3% |
| Standard | 1.5% | 1.25% | 5% |
The recommended portfolio also avoids repeating the same leg or reusing the same game across multiple suggested entries. A valid output can be SKIP or zero recommended entries.
Important payout limitation
The current Builder assumes an all-or-nothing total-return multiplier including stake. Flex / partial-payout structures where some losing-leg combinations still return money require a separate payout-state model and should not be treated as equivalent.
Publisher mode
Publisher mode separates official Bozzle records from ordinary visitor research.
- Publisher authentication is handled through Supabase Auth.
- Qualified exact-line PICKs can be locked as official Bozzfather targets.
- Publisher Model Ledger records can sync to the cloud for cross-device research and Model Lab.
- Public visitors can research and maintain local Prospects but cannot publish official targets.
- Bankroll Builder's bankroll, unit and stake suggestions are not published to Bozzfather.
Technical architecture
Engine isolation
Sport is part of the modeling boundary. NFL requests route to NFL logic; MLB requests route to MLB hitter/pitcher models. Shared UI infrastructure does not imply shared projection formulas.
Model versioning
Predictions store the model version that created them. This allows Model Lab to compare generations rather than silently overwriting history after an engine changes.
Storage & privacy
| Data | Current storage | Notes |
|---|---|---|
| My Prospects | Browser local storage | Private to the current browser profile in this version. |
| Bankroll Builder settings | Browser local storage | Bankroll, units, risk mode and entered multipliers are not shown on the public Bozzfather page. |
| Daily Scan cache | Browser session storage | Disposable performance cache; generally cleared with the browser session. |
| Local Model Ledger receipt | Compact browser local storage | Unsettled receipts are protected; older settled local rows can roll off as the browser cache grows. |
| Publisher Model Ledger | Supabase cloud | Used as the long-term official research archive when Publisher mode is active. |
| Official Bozzfather targets | Supabase cloud | Public-facing target/history data. |
Provider API credentials are intended to stay server-side in the Netlify Function environment rather than in public JavaScript.
Current limitations
- Bozzle is still accumulating out-of-sample receipts; displayed probabilities should not be treated as proven long-run edges until Model Lab supports that conclusion.
- Sportsbook / fantasy lines can move after a scan. Re-evaluate the exact line you actually see.
- Provider coverage can be incomplete, delayed or inconsistent for some markets.
- Not every available market has a supported production model. Unsupported markets are not force-scored.
- MLB lineup/probable-starter information becomes materially more useful closer to first pitch.
- Bankroll Builder does not yet model Flex / partial-payout states or trained same-game covariance.
- Bozzle does not place entries, connect to a sportsbook account or guarantee profitability.
Glossary
- Bozzle probability
- Final displayed calibrated probability for the predicted side.
- Projection
- The model's expected statistical output before comparing it with the selected line.
- MORE / LESS
- The side Bozzle estimates as more likely relative to the selected line.
- Market consensus
- Sportsbook reference formed from available book lines.
- Opening line
- The earlier/opening sportsbook reference when available.
- Line value
- How the exact line you entered compares with the current sportsbook reference from the perspective of the selected side.
- Calibration
- How closely predicted probabilities match observed frequencies over many settled predictions.
- Correlation
- Two outcomes sharing underlying causes so they are not statistically independent.
- EV
- Expected value: the model-estimated average profit/loss per unit if the same probability/payout situation could be repeated many times.
- Kelly criterion
- A bankroll-sizing formula based on probability and payout. Bozzle uses only a fraction of full Kelly and applies additional caps.
- Brier score
- Probability forecast error score; lower is better.
- Receipt
- A frozen pregame prediction stored so it can be graded later without hindsight.
Troubleshooting & FAQ
Bankroll Builder says no current scan was found.
Run today's NFL or MLB Daily Scan in the same browser session, then return to Bankroll Builder. The Builder can also fall back to current pregame Model Ledger receipts when available.
A signal is over 65% but says REVIEW.
The probability threshold and PICK gate are separate. REVIEW usually means context, market breadth/agreement, player status, lineup/start confirmation or another quality gate is incomplete.
Why did Bozzle choose LESS even though the player's recent average looks high?
The selected line, full distribution, opportunity model, market reference, recent hit pattern and calibration can all matter. The average alone is not the probability model.
Why are Hits and Total Bases treated as correlated?
A hit directly contributes to total bases, so the two outcomes share much of the same underlying event. Counting them as independent evidence would overstate sample strength and can overstate multi-pick joint probability.
Why are same-game Bankroll Builder combinations blocked?
Until Bozzle has a validated covariance model, multiplying same-game leg probabilities as if they were independent can be misleading. Blocking them is the conservative default.
Why did the public Bozzfather page show nothing?
That is valid behavior. Bozzfather only shows official published targets. If nothing earned publication, the chair stays quiet.
Will clearing browser data affect Bozzle?
It can remove local Prospects, Bankroll Builder settings and other browser-local data. Publisher cloud records and public Bozzfather history are separate.
Does a 70% model probability mean the pick will win?
No. It means the model estimates a 0.70 probability under its current assumptions and calibration. A properly calibrated 70% bucket should still lose about 30% of the time over a large sample.
Use Bozzle as a filter, not a promise.
The product is designed to reduce noise, make assumptions visible and keep an auditable record. The strongest evidence that Bozzle is useful will come from its settled out-of-sample Model Ledger—not from any single winning or losing day.