forecastscope

Finite-prefix forecasts, temporal combinations of reference experts, and sequence diagnostics in a reusable Python library. The core requires NumPy. Diagnostic implementations use optional scientific packages through forecastscope.diagnostics.

Release availability is recorded on PyPI and GitHub. Application adoption is a separate step. The extraction preserves numerical equations while adding strict public request and output validation, lint normalization and warnings for unavailable optional diagnostics. It makes no new forecast accuracy, calibration, chaos-identification or operational-performance claim.

import numpy as np
from forecastscope import combine, forecast

history = 20 + np.tile([1., 4., 2., 3.], 30)
columns = forecast("HoltWinters", history, seasonality=4, horizon=8)
allocation = combine("ValidationMix", history, seasonality=4, horizon=8)
print(columns.shape, allocation.weights, allocation.origins)

The caller supplies a contiguous observed history; raw arrays contain no timestamps, so contiguity itself is a caller precondition. Future targets belong to the caller's evaluation protocol. NaNs, infinite values, fractional horizons, invalid seasonal lags and unsupported probability grids are refused by the public entry points. The reference kernels preserve their documented short-history policies, including Holt fallback when additive seasonal smoothing lacks two cycles or has lag one. Combination experts can consequently coincide on a nonseasonal input. The five reference experts use approximate Gaussian residual intervals; ordered columns do not establish calibration. Temporal validation filters an outer-prefix-defined historical grid, and does not reconstruct the exact grid a historical service would have proposed at every earlier date.

The fixed contextual combination has an unfavorable comparison with its matching recency control in its original development study. Later validation-selected combinations have bounded point-score improvements on specified external collections, with stronger automatic-method and interval-score comparisons retained. The package does not turn those studies into a general superiority claim.

For diagnostics, install the diagnostics extra and import the topic module, for example forecastscope.diagnostics.sequence.matrix_profile. Diagnostics describe their supplied sequence and can be retrospective. They must not be computed on a future block and then treated as prefix-only model-selection inputs.

The source origin and exact input hashes are recorded in docs/extraction-source.json. The originating software is MIT licensed; its copyright and licence remain in this repository. Application cases, generator seeds, source CSVs, artifact schemas, pipeline stages and web assets are outside this library. See the extraction boundary.