cascade-fullctx-research-v16

A differentiated NumPy/SciPy full-context generator derived from the public custom-fullctx-v4 design.

Version 16 preserves the trained v15 generator and applies one targeted fix identified by its 4,096-series-per-family evaluation. Hard sensor censoring and finite-range quantization remain available to bounded families, but are no longer applied to unbounded integrated paths. Prefix-calibrated hard bounds had turned some random walks into absorbing flat lines across the forecast horizon, artificially giving persistence zero MASE.

All other v15 refinements remain: non-circular GP paths, exact Davies-Harte fGn, stateful count and intermittent demand, and conditional pulse processes.

The refinement retains hard subtypes rather than optimizing for an easy synthetic self-score: integrated paths keep genuine I(1) and I(2) branches while placing more mass on learnable persistent velocity, and 20% of OU rows retain an ultra-slow near-unit-root rate while 15% retain rapid hidden switching. GP and long-memory paths use 2L embeddings so emitted segments no longer end at an artificial circular wrap boundary.

Design

• Emits only 4096-point series, matching the 128-patch training/eval geometry and minimizing the one non-target context patch paid per series.
• Samples trend as total end-to-end excursion, so trend strength is independent of series length.
• Mixes fourteen families: trend/seasonality, structural regimes, multiplicative series, AR/integrated/nonlinear dynamics, smooth spectral GP, power-law long memory, physical sensors, seasonal counts, intermittent demand, and outliers.
• Replaces the predecessor's slow 48-pass random-Fourier GP with one batched inverse FFT and adds persistent/anti-persistent spectral paths.
• Adds a small regime-switching mean-reverting family with bounded clustered volatility, heavy-tailed innovations, transient shocks, and seasonal means.
• Executes the OU recurrence with SciPy's compiled linear filter instead of a 4095-step Python loop.
• Executes AR(1) and AR(2) recurrences with compiled SciPy filters; local microbenchmarks were 4–12× faster than scanning time in Python.
• Adds batched cadence-seasonal Poisson and gamma-mixed Poisson counts with signed trends and decaying bursts, preserving positive integer structure.
• Adds weekday/weekend interactions to a minority of count rows and piecewise spectral slopes to long-memory rows, covering calendar effects and scale-dependent roughness without another FFT.
• Gives 40% of trend-seasonal rows a low-innovation mode, teaching sharp, stable periodic reconstruction while retaining noisy seasonal coverage.
• Anchors most mixture mass on the five-seed-tested full-context core (trend/seasonal, regimes, multiplicative, AR2, and integrated paths), while retaining each newer prior at a conservative share.
• Uses published TempoPFN ablations to strengthen OU/SDE-like dynamics, spectral/long-memory paths, and transient events without letting one family dominate the corpus.
• Adds slowly modulated amplitude and phase to a minority of seasonal components; stationary cycles remain the majority. Multiplicative paths use the same full cadence bank instead of a four-period subset.
• Extends structural/event coverage with piecewise-affine regime trends, decaying shock recovery, event plateaus, and genuine held-constant runs.
• Keeps independent outliers as only 15% of the 1%-weight pulse family; most pulse rows instead use jittered cadence, seasonal event intensity, or a stable Hawkes-like recurrence whose near-future hazard responds to history.
• Applies low-rate causal censoring, finite-range quantization, and sample-and-hold artifacts to bridge clean priors to real measurement pipelines. Time reversal is limited to laws valid under reversal.
• Extends seasonality through 365/672/730-step cycles and adds a small generic physical-sensor family (smooth, bounded, pressure-like, and skewed-positive) without adopting the competitor's private-pool-shaped weather weighting.
• Generates lazy 2048-row random-family chunks, keeping every stream prefix mixed while amortizing Python dispatch. Local profiling found this about 6% faster than 1024 rows; 4096 rows regressed slightly.
• Evaluates optional seasonal components only for active rows while preserving the fixed RNG draw sequence, and caches the fixed cadence sine/cosine basis, reducing trigonometric work without narrowing the prior.
• Draws jump, shock, and heavy-tail values only where those sparse branches are active instead of allocating dense arrays whose values are mostly discarded.
• Corrects the AR(2) family to remain stationary instead of adding a 20-unit full-context drift, and scales I(2) paths by L rather than sqrt(L) so they no longer dominate I(1) paths.
• Burns in stationary AR(2) and SETAR recurrences rather than emitting their arbitrary zero-state transient.
• Calibrates standardization and measurement thresholds on the first 512 observations, preventing generated history from depending on unseen targets.
• Excludes hard range artifacts from integrated paths so an unbounded process cannot become an artificial absorbing plateau. Sign inversion and causal sample-and-hold remain available.
• Uses ForecastPFN-style expectation-centered Weibull multiplicative noise, chaotic-map burn-in with optional observation noise, and ±5% period jitter on slowly evolving seasonal components.
• Samples an actual RBF/Rational-Quadratic covariance mixture through circulant FFT embedding, rather than labeling a generic heavy-tailed frequency envelope as Rational Quadratic.
• Samples stationary fGn with beta=2H-1, then cumulatively sums selected paths into mathematically consistent fBm.
• Replaces permanent volatility jumps with bounded mean-reverting log stochastic volatility calibrated to TempoPFN's OU ranges.
• Makes intermittent demand genuinely zero-inflated with a correlated occurrence state and positive integer sizes. Count artifacts preserve integer values and causal direction.

The v13 predecessor generated a five-run median of 6.59M points/s versus 7.23M for the archived v12 baseline. The exact composite covariance and richer observation models cost about 8.8% throughput, but the candidate remains 78% above the mainnet contract's 3.7M reference in isolation. End-to-end token completion also includes model training and stream handoff. Re-measure v16 before treating those inherited numbers as current.

An end-to-end isolation run found that synchronous generation left training blocked on data for 21.9% of its wall (2.13M point-passes/s). A deterministic one-chunk producer thread now overlaps NumPy/SciPy generation with GPU work, cutting data wait to 3.9% and raising training throughput to 2.43M point-passes/s (+14.4%) on the A100. The same short contract budget then completed without a deadline hit. Cached rows reached 2.74M, confirming the remaining gap to the live L40S reference is mostly model/device throughput.

A controlled 120-second parameter screen then compared the baseline mixture with seasonal-, spectral-, and dynamics-heavy variants under the same model, pool, budget, and seeds. Dynamics-heavy won all three validation seeds, reducing mean local synthetic-pool geomean from 0.19097 to 0.18431 (3.5%; lower is better). The applied weights increase AR(2), integrated, threshold-AR, chaotic, regime-shift, and OU coverage while reducing stationary seasonal, spectral, and sparse/count families. This remains a directional local result, not a live validator verdict.

Local training result

The v10 corpus was trained under the mainnet chain.toml contract on an A100 for the full 3-hour wall. It scored 0.13679 on the 64-window local synthetic smoke pool (lower is better), improving from 0.15429 at the 30-minute heat budget, while reaching 55% of the token budget. The optimized dynamics-heavy v11 heat reached 59% (3.90B / 6.66B) and scored 0.15424. The v12 prefetch isolation test then cut data wait from 21.9% to 3.9% and raised end-to-end throughput from 2.13M to 2.43M point-passes/s. These scores are directional and are not live-validator verdicts; the A100 remains below the contract's L40S-calibrated 3.7M reference.

Research basis

• TempoPFN supplies the regime-switching OU ranges, seasonal jitter, bounded stochastic-volatility rationale, and evidence for diverse step/spike priors: https://arxiv.org/html/2510.25502v1.
• ForecastPFN motivates multiplicative Weibull noise whose center does not bias the underlying signal: https://arxiv.org/abs/2311.01933.
• Chronos KernelSynth motivates RBF and Rational-Quadratic covariance composition: https://arxiv.org/html/2403.07815v1.
• Fractional-process literature gives beta=2H-1 for fGn and beta=2H+1 for fBm: https://pmc.ncbi.nlm.nih.gov/articles/PMC3947294/.
• Intermittent-demand state-space work separates occurrence probability from positive size: https://mpra.ub.uni-muenchen.de/82487/.
• Hawkes-process literature motivates a conditional event intensity increased by recent arrivals: https://arxiv.org/abs/2405.10527.
• Heavy-tail forecasting work motivates retaining unpredictable innovations for likelihood and tail calibration rather than exact timing prediction: https://arxiv.org/abs/2106.10952.
• Proper-scoring-rule theory motivates honest predictive distributions for irreducible event uncertainty: https://doi.org/10.1198/016214506000001437.

The trend, regime-step, SETAR, physical-sensor, seasonal-count, and pulse families retain their prior broad parameter ranges where no source establishes a universal Toto2-optimal distribution. Their implementations were checked for stability and structural validity; changing every number would be false precision. Family weights also remain the locally screened dynamics-heavy mix.

Validate

python -m cascade.miner.cli verify ./generators/cascade-v16 --chain-toml chain.toml

Contract validity and CPU throughput do not establish forecasting quality. Run a production-faithful GPU A/B score against the current king before deploying this candidate.