rel-amazon: user LTV (Regression)

This runner now materializes the official rel-amazon tables through the relbench Python library and trains on the library-aligned temporal split instead of the old S3 snapshot flow.

from relbench_amazon_user_ltv import run_rel_amazon_user_ltv

df_val, df_holdout, holdout_mae, n_features, materialized, target = run_rel_amazon_user_ltv()
print(materialized)
print(target, holdout_mae, n_features, len(df_val), len(df_holdout))

Current implementation:

• examples/relbench_amazon_user_ltv.py
• examples/relbench_amazon_common.py

Current RelBench-aligned dates:

• validation cut date: 2015-10-01
• holdout cut date: 2016-01-01
• lookback start: 2008-01-01

Latest Tested Result

Latest local run from tests/data/relbench/run_reports/relbench_results.json:

• test MAE: 14.0235
• test RMSE: 48.0019
• feature count: 30
• rows: train 4708383, validation 409792, test 351885
• runtime: 1034.9s (2026-04-09T22:23:37Z)

Interactive Runner

Save API URL Run rel-amazon user LTV

Idle