Hedonix Research

Abstract

This technical note documents the production cutover, on 15 May 2026, of the H6 model ecosystem from the OLS hedonic specifications of Hedonix Working Paper No. 2026-02 to a coherent XGBoost ensemble trained on a Sun & Moon–through–Modern-Era panel of 1,733 cards (PSA-10 and PSA-9 targets) and 2,822 cards (raw target). The cutover is motivated by an empirical plafond observed during a four-round refit cycle: OLS on the broader heterogeneous-rarity panel loses approximately 15 percentage points of leave-one-set-out R² on PSA-10 relative to the narrow v3 SV-only panel. A tree-ensemble architecture on the same panel recovers 5 percentage points of the gap on PSA-10 and strictly beats the narrow v3 baseline on PSA-9 (LOSO R² 0.812 vs. 0.745, median absolute percent error 19.4% vs. 24.3%). We document the three production models, their in-sample, 5-fold cross-validation, and leave-one-set-out metrics, and the methodological rationale for accepting a stress-test LOSO penalty in exchange for production-realistic coverage. We additionally introduce a three-tier cohort system that assigns each cohort-eligible card three independent quintiles — one per grade tier — and decouples this analytic surface from the strict 648-card raw-XGB universe that backs the live Hedonix Index NAV. This note is an architectural change-log; the live-track and walk-forward analyses on the v8 architecture remain forthcoming.

Abstract

This paper updates and extends the hedonic pricing framework introduced in Hedonix Working Paper No. 2026-01 along five dimensions. We integrate PSA Population Report data and lift in-sample R² from 0.879 to 0.914. We report a parallel PSA-9 hedonic specification (n=347, R² 0.920). We expand the analytic universe from 360 to 2,635 cards across three modern eras and develop a non-parametric raw-price XGBoost model (LOSO R² 0.83). We document out-of-sample validation along four axes (random k-fold, leave-one-set-out, 180-day convergence backtest, single-anchor live track). A post-hoc multi-anchor robustness sprint stress-tests the single-anchor result and finds that cross-sectional return spread is not multi-anchor robust, while cross-sectional risk spread is robust at the 100% of-anchors level (Q5 Sharpe 2.88 vs Q1 1.60, vol 24% vs 40%). We re-position the framework as defensible cross-sectional risk management, not directional alpha generation.

Abstract

We apply hedonic pricing methodology to the secondary market for graded Pokémon trading cards, focusing on Special Illustration Rare and Illustration Rare cards from fourteen Scarlet & Violet expansion sets released between 2023 and 2026. Using transaction-level data on 360 cards, we estimate log-linear hedonic regressions in which the dependent variable is the natural logarithm of average eBay PSA 10 sale prices. The fully specified model achieves an R² of 0.879 (adjusted R² of 0.870). Raw card price exhibits a robust elasticity of approximately 0.48, with significantly steeper price transmission for Special Illustration Rares than for Illustration Rares. Top character tiers, artist identification, and a hype-divergence proxy all contribute statistically and economically significant explanatory power. The methodology described in this paper underlies the H6 Hedonic Engine, the pricing core of the Hedonix platform.