SHOPLYTICS — Cold-Start Demand Forecasting for New Fashion Products
Phase-adaptive donor-guided forecasting on the VISUELLE 2.0 SO-fore benchmark
IN PROGRESSTEAM
Overview
Final-year research project tackling the Short-Observation Forecasting (SO-fore) sub-problem of New Fashion Product Performance Forecasting. Built a weekly-native pipeline that combines a phase-adaptive multi-metric similarity engine, donor-guided LSTM transfer fine-tuning, and a BCa-bootstrap negative-transfer gate, and evaluated it on the official VISUELLE 2.0 SO-fore_2-10 test split (10,684 item-shop pairs). Reached WAPE 90.82, a paired-bootstrap statistically significant 2.17-point improvement over the strongest non-image baseline; on the low-volume tier (the SME deployment regime), the improvement reaches +9.54 WAPE. Every result reproduces from a single shell command; the repository ships a frozen test-set manifest with content hashes, 330 automated tests, and 16 named experiments documenting both successful results and surfaced research bugs (oracle-gate test-set leak caught on a 50-pair smoke run; pandas dtype-mismatch silently disabling an entire similarity branch).
Gallery
Project Details
ClientFinal-Year Project — University of Moratuwa (Supervisor: Dr. M. G. Upeksha Ganegoda)
RoleResearcher, Engineer, and Author
Timeline2025-11-30 - 2026-07-30
DurationOngoing
Technologies
PythonPyTorchOpenCLIP (ViT-B/32)ResNet-101pandasNumPyscikit-learnstatsmodelsStreamlitFastAPIpytest
Key Features
- Phase-adaptive progressive similarity engine (Contribution 1): The active set of similarity metrics adapts to the observation window length — Phase A uses metadata + image; Phase B adds demand correlation; Phase C adds seasonality alignment + volatility matching. Heterogeneous metrics fused by parameter-free rank averaging.
- Donor-guided LSTM transfer fine-tuning (Contribution 2): Weekly-native LSTM pretrained on 96,166 (product, store) pairs; per-target frozen-base head fine-tune on the top-K=20 donors selected by Contribution 1. Paired bootstrap shows the LSTM significantly improves over the donor-weighted aggregator (skill +0.013, 95% CI [+0.010, +0.016]).
- BCa-bootstrap negative-transfer gate (Contribution 3): Per-product runtime decision rule based on the lower bound of a 95% bias-corrected-and-accelerated bootstrap CI on the skill score s = 1 − WAPE_transfer / WAPE_baseline. Decision boundary at zero, no tuned threshold. Diebold-Mariano test retained as an independent diagnostic.
- Reproduced Skenderi 2022 Tier 1+2 baselines: Naive, SES, kNN time-series (k=11), and kNN+image (CLIP and ResNet-101) reproductions on the official VISUELLE 2.0 SO-fore_2-10 split, all within published tolerance. Both image variants reproduce the documented qualitative finding that motivates Contribution 1.
- Statistically rigorous evaluation: Paired bootstrap CIs (1000 resamples) on every WAPE / MAE / skill cell. Population-level skill score replaces a per-pair gate at T_obs=14d, where rolling-origin CV cannot produce enough folds for a meaningful BCa CI.
- Per-tier breakdown surfacing SME-relevant value: Tercile analysis on observed-mean weekly volume reveals the biggest win lands on low-volume pairs (n=2,337): WAPE 87.23 vs kNN-ts 96.77, a +9.54 WAPE improvement — the cold-start regime SHOPLYTICS targets.
- End-to-end reproducibility from a clean checkout: scripts/run_official_evaluation.sh reproduces the entire baseline table; scripts/module1/run_official_pipeline.py reproduces the headline method. Frozen test-set manifest with SHA256 hashes; pretrain manifests record holdout exclusion; 330 automated tests.