Forecasting the spread, honestly

negative-resultforecast

Question

Whether a fitted model forecasts the spread better than a naive persistence null.

Verdict

Pooled R2 beats the harder null (+0.0729 vs -0.0666) but per-item (equal-weighted and notional-weighted) R2 LOSES to it (-0.4959 vs -0.3538; -1.6372 vs -0.4366). Reporting only the pooled figure 'would have been selection dressed as a result.'

In plain language (3 terms)

This site's own explanation of the words above. The verdict itself is published exactly as the research index records it and is not rewritten here.

pooled: One number computed over every item's rows thrown together. A pooled figure can look strong while the per-item figures are weak, which is why both are always reported here.
null: The dull baseline a model has to beat, such as assuming tomorrow looks like today. It is chosen and published before any model runs, so it cannot be picked afterwards to make a result look better.
notional-weighted: Two ways of averaging across items. Equal-weighted counts every item the same. Notional-weighted counts the items with more money moving through them for more.

Population: 381 items, 434,492 rows.
Resolution: Daily resolution (2020-08-01 -> 2026-01-31).
Cutoff: Development span only (2020-08-01 -> 2026-01-31).
Baseline: The harder naive/persistence null from forecast-baselines.md.

Limitations

Pooled R2 beats the null but per-item (equal-weighted and notional-weighted) R2 LOSES to it -- reported on the population that matters operationally (per-item), not smoothed into the pooled figure.

Notes

REQUIRED NEGATIVE RESULT for issue #130's index acceptance criterion -- reported as negative on the population that matters operationally (per-item), not smoothed into a qualified positive via the pooled figure. docs/writeup/technical-narrative.md §5 cites this document under 'issue #48'; #48 does not itself appear inside spread-forecast.md's own text (only #42, #46, #57, #64, #66 do) -- flagged here rather than silently resolved, per issue #130's accuracy rule. Also contains the two leakage demonstrations retold in docs/writeup/failure-stories.md story 5.

Source

Citation: po-research:spread-forecast:2988c9ed4e68311d
Write-up: docs/research/spread-forecast.md
Data: docs/research/spread-forecast.json
Artifact: research-index.json v1 — sha256:cb3d976779a5cbcb4188039a8d26d5e558a2d844030a638c2f16fe5a8eb0a2fe
Build data cutoff: 2026-08-12
Issues: #42, #46, #57, #64, #66