solo-levelling-linear-regression

Solo Levelling — Linear Regression Strength Predictor

Model Description

A regularized (Ridge) multi-output linear regression that predicts next week's bench, squat, and deadlift from weekly bodyweight, calorie intake, and lift history. It also powers simple calorie-recommendation and plateau-detection heuristics. This is the baseline and template model in the four-model Solo Levelling study (linear regression, XGBoost, LSTM, Transformer).

Intended Use & Users

Personal strength forecasting and training guidance for a single athlete. Not medical or nutritional advice.

Training Data

Weekly self-logged measurements (bodyweight kg, avg daily kcal, three lift working weights). Units: kg, kcal. Chronological 70/15/15 split — no random shuffling, since this is a time series. Transformers/scalers are fit on train only to prevent leakage.

Note on this checkpoint: the published weights are trained on the repository's deterministic synthetic sample data (make sample-data, seed 42), so the repo is fully reproducible without private data. Retrain on your own logs to get a personalized model.

Features

Lagged values (1–4 wk), 3/4-week rolling means and slopes, week-over-week deltas, rate-of-change, kcal_per_kg, and total_load. All features derive from past rows (.shift(1) before rolling), never the concurrent week.

Evaluation

• MAE (kg) and RMSE per lift and overall.
• MASE vs a naive persistence forecast (next week = this week) — the primary honesty metric. MASE < 1 beats persistence.
• A CI quality gate rejects any model with MASE ≥ 1.0.

Limitations & Bias

Trained on a single individual with a small dataset (~1 row/week). It will not generalize across bodies, programs, or training styles. On data this small, this simple linear model is expected to match or beat the deep-learning variants — the LSTM/Transformer repos exist to demonstrate architecture, and only overtake at multi-user scale.

How to Use

import joblib, pandas as pd
blob = joblib.load("ridge_strength_v1.joblib")
pipeline, features = blob["pipeline"], blob["features"]
X = pd.DataFrame([{...}])[features]   # engineered feature vector
pred = pipeline.predict(X.values)     # -> [[bench, squat, deadlift]]

Reproducibility

Every run logs the git SHA, resolved config, seed, and feature list. Training run tracked in MLflow. Regenerate the exact sample data with python -m solo_linreg.data.generate.

Citation

Part of the Solo Levelling AI fitness analytics study. Code: GitHub solo-levelling-linear-regression.