Token forecast

Model and market context

Hybrid08 is our own model, trained on our BTC/ETH/SOL exchange history, with no Kronos pretrained weights. A causal per-candle tokenizer represents signed returns and candle/activity state. The native-hour transformer observes all three assets jointly and factors each future hour BTC → ETH → SOL, with return then state tokens. A small pooled LightGBM direction expert uses causal five-minute price, flow, funding and cross-asset correlation/beta features. Its tree count and probability calibration were fitted using chronological folds inside TRAIN only.

Every forecast consumes the same last 6,144 completed five-minute candles per asset. They form 512 nonoverlapping hourly candles and 511 hourly observations; the neural encoder uses 128 hours after causal feature warmup. Price tokens use the known preceding 48-hour RMS return scale. Later scenarios update volatility and context from their own generated past, never actual future candles. Funding is aligned causally. Open interest stays disabled because TRAIN did not contain usable OI; live availability cannot enable an untrained input.

The direction expert reweights 256 genuine joint neural scenarios using iterative proportional fitting. It does not edit predicted prices, flip asset directions or force ETH/SOL to follow BTC. Unsupported targets shrink toward the original scenario mass or fall back to uniform weights, with a minimum effective sample count. All reported forecast scores use achieved weights and actual weighted price quantiles, rather than substituting classifier accuracy. Different near-50% marginal probabilities can give opposite small median moves even when most joint scenarios move together.

How the forecast is drawn

• One native candle spans one actual hour starting at the common five-minute input cutoff; its phase need not equal the exchange’s UTC-hour open. There are 24 hourly scenarios. With Candles selected, whole-hour multiples such as 1h, 2h, 4h and 1D aggregate consecutive model candles using first open, maximum high, minimum low and last close. Sub-hour chart intervals, including custom minute intervals, slice a deterministic illustrative path: open → low → high → close for a rising hour, or open → high → low → close for a falling hour, with extrema at 20 and 40 minutes. A piecewise linear reference supplies display opens/closes and native extrema; additional deterministic wicks are bounded by the overlapping hourly high/low. This preserves the hourly closing path and total range, without predicting when intrahour moves will happen. The legend explicitly labels interpolated candles and illustrative intrahour movement. The rolling display keeps the model’s actual origin; it does not relabel its values as exchange UTC buckets. Each body occupies one selected chart interval and uses the traded candle width and shared price transform. A trailing partial interval connects the last illustrative close to the remaining hourly endpoints as a dashed line. Larger intervals that are not whole-hour multiples, or a horizon with no complete display candle, use the hourly close line. Switching interval rebuilds only the display; model outputs, accuracy scores, alerts and model-led plans continue to use genuine hourly predictions. The hourly p10–p90 close range remains separate from illustrative candle wicks. The regular exchange time grid is preserved without inserting off-grid native-series timestamps.
• The forming traded candle owns its entire selected chart interval. Forecast bodies whose open precedes its close are omitted, including when new exchange bars arrive; no prediction timestamps are moved to create a gap. The Forecast candles marker identifies the first visible future body. The underlying hourly forecast origin and endpoint times stay unchanged. If no complete future display candle fits, show the close line instead. Sub-hour illustrative wicks extend beyond their bodies, capped by the high/low of every overlapping native hour. The base excursion is 0.3 × hourly envelope range × sqrt(selected seconds / 3600), with fixed four-candle upper factors [0.75, 1.15, 0.9, 1.05] and lower factors [1.1, 0.8, 1.2, 0.7]. Source path extrema are retained. This deterministic display template changes neither closing prices nor model probabilities and does not claim to predict fine-resolution highs or lows.
• For new preferences, the chart defaults to the one-hour median line anchored to the latest traded price. Existing style and horizon choices remain available. Live anchoring is automatic; a saved legacy model-price preference is ignored. The shaded range joins native p10 and p90 close quantiles; intermediate edges are visual joins, not extra predictions. The native hourly median body is continuous from the observed anchor, though the first visible future display body can start later. U and D are weighted medians of each scenario’s log upper/lower within-candle excursions. Native high = max(open, close) × exp(U), low = min(open, close) × exp(−D). These illustrative wicks are not prediction intervals and the displayed candle need not be one sampled trajectory.
• Live anchoring always multiplies the original displayed candle prices and close-range prices by the same latest-traded-close / first-remaining-model-open factor. Every exchange-price update derives a fresh display from the untouched model prices, so repeated updates cannot accumulate drift. The first range point joins the live price at that remaining hour’s original open; elapsed range points are omitted, while future prediction timestamps remain unchanged. Candles, the median line, the uncertainty band and model-led plan references therefore use the same price basis. This is a display rebase, not a new inference: archived model prices, scores, quantiles and probability mass remain unchanged, and fresh scenarios still arrive each completed five-minute input close. Gain/loss colors retain their normal meanings. The summary calls probabilities between 40% and 60% No clear tilt, rather than presenting small opposite medians as confident opposing calls.
• One shared authenticated worker generates all three assets at a common cutoff, refreshed each completed five-minute input close. It rechecks consumed final candle fields after inference and rejects changed input or anything older than one refresh cycle. A boundary crossing preserves the original cutoff and returns a labelled previous-cycle result while the latest cycle computes. Model weights, expert/calibration, qualification evidence, source files and runtime identity are checked. The server atomically caches the joint bundle and deduplicates same-cutoff requests across market switches. The previous cycle may remain labelled while a replacement computes; unavailable or stale output cannot trigger a model alert.
• The retained snapshot component, outside the default workspace, is an explicitly offline view from the same model/version and native interval. Its selected horizon changes both metrics and the plotted range. Snapshot freshness is measured from the input close, not the preceding five-minute candle-open label. Alerts and model-led planning use the qualified one-hour horizon; 4h and 24h remain labelled exploratory and cannot create model-agreement alerts or model-led plans.

Measured accuracy and limits

• On 1,007 development timestamps and 3,021 correlated asset outcomes, Hybrid08 direction accuracy was 53.33% with a UTC-day block 95% interval of 50.89–55.73%. Corrected v1 scored 50.22%; price MAE fell from 49.80 to 33.73 basis points, versus no-change 33.84. Repeated development selection makes this an optimistic estimate. All 1,007 outputs were then reproduced through the bounded-past serving adapter with identical directions.
• With learned settings frozen, complete 42-day reused historical confirmation scored 52.66% [50.38–55.08], versus v1 50.78%. BTC scored 50.84%, ETH 54.12%, SOL 53.03%; individual intervals remain wide, especially BTC. MAE was 32.02bps versus v1 46.81 and no-change 32.12. Nominal 80% range coverage was 77.19%. The original pooled one-hour checks passed. This holdout had already been inspected for older models, so it is historical confirmation, not a pristine final test or live accuracy guarantee.
• The previously unscored later snapshot contained only 23 timestamps/69 correlated outcomes: 53.62% direction versus v1 44.93%, but below momentum’s 55.07%; SOL also trailed v1 on that short span. It is supplemental retrospective evidence, never live-issued accuracy. Do not infer a reliable confidence interval from roughly one day. Actual future-issued research forecasts preserve issue time and input cutoff; their outcomes are recorded only after exchange maturity.
• Longer development tests used genuine recursive hourly scenarios and nonoverlapping outcomes. Four hours scored 49.27% across 753 asset outcomes and failed checks; 24 hours scored 57.72% across only 123 outcomes with an interval of 43.09–71.54%. Directional skill is not established for either. No longer-horizon probability calibration was fitted; the original first-hour scenario weights persist over whole trajectories.
• Pinned Kronos Mini scored 53.26% [51.44–55.08] on the complete development grid and 52.27% [49.87–54.77] on the same 42-day historical confirmation used for ours (52.66%). The paired confirmation direction difference is inconclusive: +0.40 percentage points [−2.08,+3.10] for ours; price MAE is lower at 32.02 versus 36.96bps. Mini’s earlier sparse 58.33% result did not persist. Kronos’s paper headline is a relative RankIC improvement, not 93% correct market direction. Neither scenario mass nor directional accuracy is the probability of a profitable trade. Training stays offline. The separate live forecast record captures new API and scheduled-worker issues and scores exact matured exchange closes; it never imports these historical results. Continuous coverage requires the operator to run the forecast-tracking worker.