final: tier C validation counts aligned with round-3 labels
3 changed files +2 −2
modified
paper/main.pdf
+0 −0
Binary file not shown.
modified
paper/sections/data.tex
+1 −1
@@ -66,7 +66,7 @@ We match listings across banners in tiers of decreasing confidence. Tier~A is ex | ||
| 66 | 66 | |
| 67 | 67 | The thresholds were not chosen by eye. All 36{,}700 Metro$\times$Super~C candidate pairs sharing a brand and a package were scored and labelled by whether they share a UPC (4{,}328 positives). Token-set similarity---the scorer used by Food-Ka's own matcher and by a first version of ours---scores a name that is a subset of another at 1.0 (``beurre végétal salé'' matches ``beurre végétal non salé''); its precision at threshold 0.90 is far lower (0.61 on a first calibration without multipack-aware package keys). Identical token sets achieve precision 0.932 and recall 0.982; token-sort similarity $\ge 0.90$ cumulatively 0.870 and 0.987. The residual 6.8\% of identical-text pairs with different UPCs are packaging variants invisible in text, so 0.93 is a lower bound on the precision of tier~B relative to what a reader of the two listings could tell apart. |
| 68 | 68 | |
| 69 | −Figure~\ref{fig:matching}B and Table~\ref{tab:validation} report the validation on a stratified random sample of 60 pairs per tier. Each pair was labelled by Claude (claude-opus-5-5) as a second annotator and then reviewed, pair by pair, by the research operator; the two agreed on all 240 pairs after review. Tiers A and B are correct in 60/60 cases each, tier~D in 59/60 (one unsure), tier~C in only 27/60: the token-sort rule admits flavour, roast, fat-content and colour variants (``douces'' vs.\ ``piquantes'', ``torréfaction moyenne'' vs.\ ``foncée''). The preregistration therefore fixes tiers A+B as the primary specification and uses C and D only in sensitivity analyses. The tiers yield 4{,}413 (A), 4{,}088 (B), 714 (C) and 102 (D) cross-banner pairs, and 5{,}909 groups under A+B covering 13{,}088 listings; the median group spans two banners and the largest seven. Against Food-Ka's own 2{,}033 linked pairs, our tiers recover 42.5\%---the rest are size-less or multi-offer circular links that we exclude by design. | |
| 69 | +Figure~\ref{fig:matching}B and Table~\ref{tab:validation} report the validation on a stratified random sample of 60 pairs per tier. Each pair was labelled by Claude (claude-opus-5-5) as a second annotator and then reviewed, pair by pair, by the research operator; the two agreed on all 240 pairs after review. Tiers A and B are correct in 60/60 cases each, tier~D in 59/60 (one unsure), tier~C in only 25/60 (five unsure): the token-sort rule admits flavour, roast, fat-content and colour variants (``douces'' vs.\ ``piquantes'', ``torréfaction moyenne'' vs.\ ``foncée''). The preregistration therefore fixes tiers A+B as the primary specification and uses C and D only in sensitivity analyses. The tiers yield 4{,}413 (A), 4{,}088 (B), 714 (C) and 102 (D) cross-banner pairs, and 5{,}909 groups under A+B covering 13{,}088 listings; the median group spans two banners and the largest seven. Against Food-Ka's own 2{,}033 linked pairs, our tiers recover 42.5\%---the rest are size-less or multi-offer circular links that we exclude by design. | |
| 70 | 70 | |
| 71 | 71 | \begin{figure}[htbp]\centering |
| 72 | 72 | \includegraphics[width=\textwidth]{figures/fig_matching.pdf} |
modified
paper/sections/robustness.tex
+1 −1
@@ -8,6 +8,6 @@ The category-level sign tests that the median gap exceeds 5\% reject for eight o | ||
| 8 | 8 | |
| 9 | 9 | \paragraph{Falsification tests that passed.} The AR(1) placebo (permuted banner labels) returns 0.004; the H3 placebo at random dates returns $+0.0003$; single-store shops show 4.7\% multi-price weeks against 22.8\% in anti-bot feeds; the H4 leads for all-rival and same-company rivals are indistinguishable from zero. |
| 10 | 10 | |
| 11 | −\paragraph{Tests that failed or were inconclusive.} (i) The cross-company lead placebo in Model~C fails ($-0.108$, s.e.\ 0.048): with 1{,}784 observations and common weekly shocks, the descriptive lag coefficients for other-company rivals cannot be read as responses. (ii) The event-study pretrend test rejects; the decomposition by feed type attributes most of the pre-sale elevation to store-mixture noise, leaving a small exploratory residual. (iii) The market-level H5 gap is not different from zero over eleven weeks; the hypothesis holds only at the banner level. (iv) Tier~C matching failed validation (precision 0.45) and was removed from the primary specification; a first version of the matcher based on token-set similarity was rejected outright. (v) The preregistered basket rule left only seven items common to all wide-coverage banners and had to be amended (A-001). (vi) The comparison with Statistics Canada rests on two monthly observations and is inconclusive. We regard reporting these as part of the result. | |
| 11 | +\paragraph{Tests that failed or were inconclusive.} (i) The cross-company lead placebo in Model~C fails ($-0.108$, s.e.\ 0.048): with 1{,}784 observations and common weekly shocks, the descriptive lag coefficients for other-company rivals cannot be read as responses. (ii) The event-study pretrend test rejects; the decomposition by feed type attributes most of the pre-sale elevation to store-mixture noise, leaving a small exploratory residual. (iii) The market-level H5 gap is not different from zero over eleven weeks; the hypothesis holds only at the banner level. (iv) Tier~C matching failed validation (precision 0.42) and was removed from the primary specification; a first version of the matcher based on token-set similarity was rejected outright. (v) The preregistered basket rule left only seven items common to all wide-coverage banners and had to be amended (A-001). (vi) The comparison with Statistics Canada rests on two monthly observations and is inconclusive. We regard reporting these as part of the result. | |
| 12 | 12 | |
| 13 | 13 | \paragraph{Out-of-time validation.} Model~F's rolling-origin design is the only strictly out-of-time evidence in the paper; the promotion-state forecasts (AUC 0.94) are the component of the architecture that generalises forward within the window. |
| 14 | 14 | |