cascade-fullctx-spectral-v13

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 fifteen families: trend/seasonality, structural regimes, multiplicative series, AR/ARIMA-integrated/nonlinear dynamics, RBF/RQ spectral GP, power-law long memory, physical sensors, seasonal counts, intermittent demand, outliers, and weekly retail/load demand.
• v13: adds weekly_demand (~8%), ARIMA-style correlated increments, 2L RBF/RQ spectral embedding (no wrap artifact), calendar seasonal_focus, heteroskedastic residuals, AR2/integrated seasonal overlays, tent-map chaos, prefix-causal measurement artifacts, forecastable pulse timing, and a fixed-length emit fast path.
• 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.
• Applies low-rate reversal, censoring, quantization, and sample-and-hold artifacts to bridge clean priors to real measurement pipelines.
• 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.

Local training result

Validate