HIGHFASHION

HIGHFASHION Runway -> Hype -> Equity HIGHFASHION turns runway-era signals into measurable forecasts. It tracks luxury houses and their parent groups. It models attention, narrative tone, and market reaction as one system. Made by Moyosore Ogunjobi.

Not financial advice.

About HIGHFASHION is a multitask temporal modeling project built around the luxury sector. The goal is to forecast house-level attention and parent-company reaction using signals that matter in both fashion and finance.

The model forecasts:

BrandHeat as a 7-day ahead regression target ParentEquityMove as a 5-trading-day ahead 3-class direction target ShockDay as a binary probability for sudden attention spikes The default entity universe includes Prada, Miu Miu, Louis Vuitton, Bottega Veneta, Maison Margiela, and TOM FORD, with a configurable expansion set for other major houses.

Data The dataset is built from public APIs with local caching for reproducibility. Raw responses are saved so runs can be inspected and repeated.

Sources:

Wikimedia Pageviews API for attention GDELT for media volume and tone Market data for parent tickers Cache layout:

data/raw/pageviews/{entity}/{YYYY-MM}.json data/raw/gdelt/{entity}/{YYYY-MM}.json data/raw/market/{ticker}.parquet Processed dataset:

data/processed/highfashion_dataset.parquet Modeling Approach The project treats luxury houses as part of an ownership graph rather than isolated brands. Each house is linked to a parent group and, where available, a public ticker.

The feature set includes:

pageviews GDELT volume GDELT tone OHLCV and market-cap context fashion-week seasonality indicators cyclical calendar features optional design motif proxy Targets are built with strict time ordering. Normalization is fit on train only. Walk-forward validation uses the earliest 70% of dates for training, the next 15% for validation, and the final 15% for test.

How To Run Build the dataset:

python -m highfashion.build_dataset
--data_dir data
--output_path data/processed/highfashion_dataset.parquet Train the model:

python -m highfashion.train
--data_path data/processed/highfashion_dataset.parquet
--output_dir artifacts/highfashion Run inference:

python -m highfashion.infer
--model_dir artifacts/highfashion Train and push to the Hub:

python -m highfashion.train
--data_path data/processed/highfashion_dataset.parquet
--output_dir artifacts/highfashion
--push_to_hub
--hub_model_id m0yosore/HIGHFASHION Outputs Training writes:

metrics.json config.json splits.json label_map.json scaler.json model.pt The test report includes:

BrandHeat: MAE, RMSE, Spearman EquityMove: accuracy, macro-F1, ECE, confusion matrix ShockDay: AUROC, AUPRC, Brier Information coefficient for signal quality Tests The repository includes tests for:

walk-forward split ordering train-only scaler behavior dataset window shapes model forward-pass output shapes CI is defined in:

.github/workflows/ci.yml Limitations This is a research-grade baseline, not an execution engine. API coverage, naming ambiguity, and sparse data for private groups can affect feature quality. Private parents such as OTB remain heat-only because no public market series exists. The optional multimodal branch is not enabled by default in this baseline.

Conclusion HIGHFASHION treats luxury like a real industry, not a toy dataset. It ships leakage-safe splits, serious metrics, and reproducible artifacts. It is built for competition settings and research-grade comparisons.