Beta · Probabilistic ForecastsBeta forecasts are probabilistic estimates. Availability depends on current data and model support.
Loading market data
Market intelligence
How Beta forecasts are made, recorded and verified.
Forecasts are provided where data supports them, and validation continues in the background.
Eligible active assets may receive 1H and 24H Beta forecasts. Market Status describes the observed market; Forecast Status describes model availability. Limited means some model outputs are supported while others remain unavailable. Price estimates can be shown without a direction probability. 7D is Coming Later.
Official Hypixel quotes must be at most 15 minutes old, as must the required feature snapshot. We require at least 2,880 distinct valid hourly observations within 150 days (80% coverage), complete finite model features, passing integrity checks, positive two-sided quantities and asset-specific evidence for the observed quote conditions. Liquidity is assessed separately from forecast reliability for individually supported assets; a wide spread can be shown with a trading-condition warning. Assets without this additional evidence retain the earlier 60/100 and 5% safeguards. Each asset and horizon needs at least 200 historical evaluation observations. The short-input 24H price model also requires 25 consecutive hourly observations covering the current hour and the previous 24 hours. Insufficient data produces Forecast Unavailable, with a reason.
P(up) estimates the chance that the future market midpoint exceeds the reference midpoint recorded at issuance. P(down) is its complement and includes unchanged prices. A 63% probability does not imply a 63% price increase. Probability describes direction; median move describes the center of a price distribution. Neither is an instruction to trade.
The 24H price candidate passed a historical development screen using complete data windows. Its median and interval come from log-price quantile models, with interval adjustment fixed on separate calibration data. The 25-observation policy passed the aggregate historical screen and matched full-history inputs in a real-data equivalence check. Expanded asset coverage is checked using denser historical evaluation records, including non-overlapping 24-hour outcomes. Each admitted asset needs at least 200 observations across 30 dates and two months, at least 20 non-overlapping outcomes, and at least 80% interval coverage in both sets. These forecasts remain Limited Beta with Low confidence. This is development evidence, not independent final validation or a per-asset accuracy guarantee. When enabled, only eligible assets receive its outputs. The 1H price model has not passed its screen. Direction probabilities remain unavailable where a supported probability model cannot run. No trading model is promoted by this feature. An expected-value regressor is not presented as a median. Executable-return quantiles from the separate shadow study are not converted into price ranges.
Some assets use a separately labelled basic forecast: their historical 24-hour log-return distribution is applied to the current quote reference. Its median and 5th/95th percentiles are fixed from December 2025 calibration observations and evaluated on January–March 2026 data. Each admitted basic asset needs sufficient calibration and evaluation support, at least 80% overall, monthly and non-overlapping interval coverage, useful interval width, and median error no more than 5% worse than a no-change benchmark. A separately evaluated variant scales the frozen historical distribution by recent 24-hour volatility, with a floor fixed from calibration data. A price-unchanged baseline uses the current reference as its median and historical return quantiles with their median removed; it claims no directional advantage. Additional variants center the historical residual distribution on the past 24-hour average price, with or without volatility scaling. The target remains the future quote midpoint. Basic forecasts are withheld when their model-derived total interval width exceeds 25% of the reference price; intervals are never clipped to fit. These are reference forecasts, not a claim of predictive advantage. Failed candidates are not published.
Some otherwise supported assets have wider uncertainty. These are shown separately as Broad reference · Low precision, allowing total interval width up to 100% of the original reference price. They must meet the same minimum history, integrity, error and coverage checks as basic forecasts, with the width condition applied using only information available at issuance. This is a broader product uncertainty limit, not evidence that the model became more accurate. The full range matters more than the median alone. Candidate methods and tiers were screened on previously inspected development data; this is not independent final validation. Live results are recorded separately by tier.
90% is the target interval coverage, not a guaranteed probability for an individual forecast. Conservative over-coverage does not by itself make a forecast unusable; width and interval error are assessed as well. The expanded Crimson Essence and Enchanted Wheat policies use previously inspected development evidence, not a new independent test. Current prices are midpoints of Hypixel quote summaries rather than guaranteed execution prices. Large spreads and limited liquidity remain visible even when a forecast is available.
Every issued Beta forecast is committed to the existing append-only ledger with its model fingerprint, reference quote, feature snapshot and eligibility evidence before display. At expiry we use the first valid official quote within the following 15 minutes, with both observation and arrival no later than the verification time. The observed midpoint and validation results are stored separately. Missing prices remain unresolved; forecasts are never rewritten or retrospectively backfilled.
Direction correctness follows the most likely probability event when present; unchanged belongs to P(down). For price-only forecasts, direction is checked from the median move. Brier score is calculated only for forecasts with probabilities and measures probability error, and calibration compares forecast probabilities with observed frequencies. Median price error and interval coverage are reported only when those model outputs exist. Historical sample support is not a guarantee of accuracy; overlapping outcomes are not independent observations.
The September 14–October 14 shadow validation protocol and its artifacts remain frozen and separate from Beta product validation. Current shadow outcomes are not used for model selection, calibration, thresholds or tuning. Beta issuance does not run Final Training or approve rejected trading strategies.
Beta forecasts are probabilistic estimates. Market changes, data gaps and limited model support can make a forecast unavailable or inaccurate. Confidence is currently Low while forward evidence accumulates. Users decide how to interpret the information; forecasts do not guarantee profitable outcomes.