Methodology
A chain of evidence with visible boundaries.
The portfolio separates model selection, validation decisions, official evaluation, finance context, and presentation. PASS means the contract and evidence reconcile—not that a model is commercially superior.
01 · Technical spine
From public source to frozen claim.
- 01
Provenance and checksums
Source roles, redistribution decisions, regular-file checks, SHA-256 anchors, and immutable receipts fail closed before evidence can be accepted.
- 02
Hierarchy
M5 results preserve the official 12-level hierarchy. WRMSSE is measured before aggregation, with validated weights, scale rules, and reconciliation fixtures.
- 03
Model selection
Seasonal Naive S7 and Candidate C were compared across four historical folds. The selected candidate recipe was frozen before validation and official evaluation.
- 04
Freeze A / Freeze B
Freeze A protects the validation replay. Freeze B V12 fixes the final training material before the isolated official-evaluation scorer opens d_1942–d_1969.
- 05
Validation and evaluation
Decision and inventory evidence use d_1914–d_1941. Official WRMSSE uses d_1942–d_1969. The windows answer different questions and are never pooled.
- 06
WRMSSE
RMSSE scale, dollar-sales weights, 12 levels, zero-history handling, and reference fixtures are fixed in the technical contract.
- 07
Decision loss
Three fixed forecasts are scored at item-store-day grain under six normalized under:over ratios; quantile recalibration is a separate, leakage-safe experiment.
- 08
Inventory exposure
A finite-horizon lost-sales simulation applies 27 predeclared policies, empirical safety buffers, and safe total/store/category aggregates. It is a proxy, not observed Walmart inventory.
- 09
SEC / XBRL
Company Facts are normalized through a concept map, canonical annual periods, accession lineage, and accounting checks. Free cash flow is derived as operating cash flow minus CAPEX.
- 10
FRED context
CPI, unemployment, Fed Funds, and retail sales are frozen latest-revised observations used descriptively. No vintage, regime, correlation, or causal analysis is claimed in v1.0.
02 · Public evidence
Open the receipts behind the claims.
Official-evaluation disclosure
Seasonal Naive S7 remained the winner.
Post-competition local reproduction; not a leaderboard submission. Labels are public today. A pre-registration availability audit is disclosed, and no evaluation values or metrics were used for model selection before Freeze B.
Candidate C did not beat the frozen baseline. Technical PASS establishes evidence integrity, not superiority.
03 · Non-intended use
What this work cannot support.
- Normalized decision and inventory costs are not USD or Walmart internal costs.
- Observed M5 sales may censor demand; inventory, orders, lead times, and policies are simulated.
- M5 estimated retail revenue is not equivalent to Walmart corporate accounts.
- SEC and FRED context do not explain the M5 result and support no causal or investment conclusion.
- No artifact is production, deployment, operational, optimality, leaderboard, or forward-guidance evidence.