Synthetic Intermittent Demand Benchmark

 Synthetic Spare-Parts Demand Benchmark (v1)

    A fidelity-certified synthetic benchmark for decision-aware forecasting of
    intermittent spare-parts demand, accompanying the paper "When Forecast
    Accuracy Hurts Service: A Cross-Family Benchmark and Bias-Direction
    Diagnostic for Intermittent Spare-Parts Demand."

    There is no public, decision-aware benchmark for genuinely intermittent
    spare-parts demand — existing forecasting benchmarks (M4/M5) are
    structurally smooth, and the industrial panels where intermittency bites are
    confidential. This dataset removes that barrier: it is a masked synthetic
    surrogate of a confidential industrial spare-parts panel, generated so that
    each material is regenerated until it passes a per-material fidelity audit
    (Syntetos–Boylan class, zero fraction, ADI, CV², zero-run distribution), and
    certified against the real data two levels deeper than summary moments — at
    the level of fitted marginal distributions (family-mix total-variation
    distance 0.044) and cross-material copula dependence (copula-family TVD
    0.081).

    Contents: 5,000 synthetic materials × 60 monthly periods, plus a multi-item
    order log and MAP unit prices; per-material demand statistics and
    fidelity-audit records; the generator and a reproduction guide. Preserves
    ~70% monthly zeros, ~90% intermittent/lumpy, right-skewed nonzero sizes,
    multi-item orders, and ~20% cold-start.

    Intended uses: decision-aware forecasting evaluation (accuracy and
    order-level service under a policy), foundation-model adaptation on sparse
    demand, intermittent-demand inventory-policy design under realistic (light)
    cross-material dependence, and bias-direction diagnostics.

    Limitations: one industrial domain, monthly granularity, 60-month window,
    light dependence regime; tail-dependence estimates noisy at 60 periods.

    No real identifiers, prices, or schema are released. License CC-BY-4.0.