Overview: This project builds a performance prediction and uncertainty quantification system for Greek raw powerlifters using the OpenPowerlifting dataset. The goal is not simply to minimise prediction error, but to build a pipeline that is honest about what the available features can and cannot predict — and to quantify that uncertainty explicitly for each athlete profile. The project is structured around eight phases, three modelling approaches, and a strength standards tool that a coach or athlete can run interactively.
Status: All eight phases complete.
Data: OpenPowerlifting, 2026-08-01 release (CC BY 4.0). Fetch instructions below — the raw CSV is not committed (> 700 MB).
Looking for a plain-language explanation instead? See
ATHLETES.md — no formulas, just what the tool tells you and
how to use it.
Greek raw lifters perform comparably to their regional peers — once the data is cleaned properly. The raw data mixed bench-only, push-pull, and full-power meets into a single "Total" column, producing a spurious bimodal distribution. After restricting to full-power (SBD) competitions, Greece sits almost on top of Southern Europe and the Balkans on every lift:
Age's effect on performance is genuinely non-linear — confirmed, not just assumed. A linear model can only fit age as a straight line; SHAP values extracted from the tuned gradient-boosted model show the actual shape: a steep rise through the mid-20s, a plateau through the early 30s, and a consistent decline thereafter.
The uncertainty intervals are honest, not just claimed. A calibration curve compares the coverage a model claims (x-axis) against what it actually delivers on held-out data (y-axis). A well-calibrated model hugs the diagonal — which is what the age-augmented BayesianRidge model does here, across the full range of confidence levels.
.
├── greek_powerlifting_analysis.ipynb ← main notebook
├── README.md ← this file (technical)
├── ATHLETES.md ← plain-language explainer
├── plots/
│ ├── phase1/ ← EDA figures
│ ├── phase2/ ← preprocessing figures
│ ├── phase3/ ← regression figures
│ ├── phase4/ ← evaluation figures
│ ├── phase5/ ← boosting & SHAP figures
│ ├── phase6/ ← Bayesian calibration figures
│ └── phase8/ ← strength standards tool output
└── data/
└── (raw CSV not committed — see below)
Goal: define a clean, well-characterised modelling cohort and understand the data's structure.
Motivation: the raw OpenPowerlifting data interleaves event types (bench-only, push-pull, full-power) in a single TotalKg column. Without understanding and correcting for this, the whole model would be built on a confounded quantity.
Approach: a seven-step filtering funnel (Greek → raw equipment → SBD full-power → valid total → no fourth attempts → 2015 onwards → deduplicated), followed by comparative analysis against Southern European and Balkan peers on all four targets.
Goal: turn the cohort into leakage-free, well-conditioned feature matrices ready for modelling.
Motivation: the scoring columns (Dots/Wilks/Goodlift) are deterministic functions of TotalKg and bodyweight — they are the target repackaged and must be excluded. Federation composition and age missingness require principled decisions before any model sees the data.
Approach: baselines first (predict-mean, per-sex mean, per-sex bodyweight regression — the floors every model must beat), then feature construction, train-only scaling, and a diagnostic suite (leakage audit, condition number check).
Goal: implement linear regression from first principles and validate it.
Motivation: implementing OLS, gradient descent, Ridge, and Lasso from scratch demonstrates the mathematics is understood, not just called; validating against sklearn confirms correctness.
Approach: normal equation via np.linalg.lstsq, then batch gradient
descent verified to converge to the same solution (< 0.001 kg difference),
then Ridge and Lasso from scratch, then sklearn validation confirming all
implementations match to four decimal places. Ridge does not improve over OLS
(condition number 3.3 — well-conditioned matrix); Lasso retains all features
with one exception: age-augmented TotalKg, where Lasso finds a real optimum
at λ = 1.79 and improves RMSE by 0.11 kg. Stratified and unstratified 5-fold
CV confirm the single-split estimates are stable.
Goal: move beyond a single held-out RMSE and determine where the models succeed, where they fail, and why — across demographic subgroups, across targets, and across the bias-variance spectrum.
Motivation: an aggregate RMSE hides subgroup-level failure. A model can look reasonable overall while explaining almost no variance for a specific demographic, and a single train/test split cannot distinguish underfitting from overfitting.
Approach and findings:
- Sex-stratified performance: RMSE is lower for female lifters across every target (smaller absolute totals), but R² is markedly worse and negative for several base-feature targets — the model explains comparatively little of the variance within the female subgroup, even though absolute errors look reasonable.
- Age-category performance: the Masters subgroup is poorly explained under both feature sets, consistent with small sample size (33–35 rows) rather than a missing feature.
- Bias-variance diagnosis: training RMSE and CV RMSE are close across every target and feature variant — the model is in a high-bias, not high-variance, regime.
- Residual correlation: squat-deadlift residuals correlate most strongly (0.80), consistent with the model's errors being shared across lifts rather than independent per lift.
Goal: explore whether a model with more functional flexibility than a linear one can improve on Phase 3–4's classical results.
Motivation: Phase 4 established the linear models were underfitting. A specific mechanism was hypothesised: the relationship between age and performance is plausibly non-monotonic — a shape a linear model can only fit as a straight line, but a tree-based model can capture via threshold splits.
Approach and findings:
- Baseline (untuned) XGBoost vs. OLS/Ridge: untuned boosting already beat the linear models on every target, with the largest gains concentrated in the age-augmented feature set (TotalKg −6.67 kg vs. OLS).
- Optuna hyperparameter tuning (100 trials per target/variant, TPE sampler) widened these gains further. Disclosed caveat: the search optimises against the same CV folds used to report results — a mild leakage likely making the reported improvement a modest overestimate.
- SHAP analysis confirmed the age-nonlinearity hypothesis directly (see figure above): Age ranks third in importance across all four targets, and SHAP dependence plots show the predicted shape explicitly.
Goal: move beyond point predictions and quantify per-prediction uncertainty. Standard Ridge gives a single number with no indication of confidence; BayesianRidge gives a mean plus a calibrated interval.
Motivation: Phase 5 produced a model that predicts well but still returns only a point estimate per athlete. Useful interpretation needs an honest sense of how much to trust that number.
Approach and findings:
- MAP-Ridge identity verified numerically: at matched λ, BayesianRidge and Ridge coefficients agree to ~1e-12 — BayesianRidge changes nothing about the point prediction, only adds calibrated uncertainty around it.
- Empirical coverage at four nominal levels (50/68/80/95%), across all eight target/variant combinations, tracked nominal levels closely (within 1–5 percentage points) — the intervals are genuinely well-calibrated on unseen data (see figure above).
Goal: give a rigorous, calibrated notion of uncertainty to the model actually established as best — tuned XGBoost — since Phase 6's BayesianRidge was built on Ridge, a model Phase 5 had already shown was beaten on every target.
Motivation: conformal prediction is model-agnostic and distribution-free, requiring no analytic posterior or Gaussian error assumption — it wraps directly around tuned XGBoost's predictions using only empirical calibration residuals.
Approach and findings:
-
Split-conformal prediction wrapped around tuned XGBoost for all eight
target/variant combinations, using the exact finite-sample quantile
correction
$k = \lceil(n_{\text{cal}}+1)(1-\alpha)\rceil$ , calibrated on a fresh 50/50 stratified split of each existing test set. - A systematic male/female coverage asymmetry emerged across all eight combinations, with no exceptions: female coverage exceeded male at every confidence level — driven by a single pooled calibration quantile dominated by the majority (male) group's error scale. Bench showed the widest gap of the three lifts (34.7 points at the 50% level).
- The same asymmetry was independently confirmed under BayesianRidge, establishing it as a property of the dataset's sex-scale heterogeneity, not an artifact of either uncertainty method.
- BayesianRidge vs. conformal-XGBoost comparison: conformal intervals were narrower in 30 of 32 target/variant/confidence-level combinations, with comparable or better coverage — most plausibly because XGBoost's residuals are simply better than Ridge's.
Goal: turn the finished pipeline into something an athlete or coach can actually use — bodyweight, sex, age, and federation in; a calibrated total prediction out.
Motivation: everything built in Phases 1–7 predicts well and knows how uncertain it is, but none of it was usable outside a notebook cell full of model objects. This phase packages the best model (tuned, age-augmented XGBoost) and its calibrated uncertainty (Phase 7's conformal intervals) into a single, editable, self-contained cell.
Approach and findings:
- Two independent predictions — one direct (age-augmented TotalKg model), one built by summing separate squat/bench/deadlift predictions — as a built-in consistency check on the model itself. Only the direct prediction carries a calibrated interval; the summed total is reported as a point-estimate cross-check only, since combining three positively correlated conformal intervals (squat–deadlift residual correlation 0.80, per Phase 4) is not a simple sum of their individual widths.
- The interval is reported at the 95% conformal level, with a plain- language note that the range was empirically checked against held-out data, not just asserted.
- A cohort-context plot places the prediction against the full
same-sex
TotalKgdistribution, answering the question the raw number alone can't: is this an average result, or an exceptional one. - Validated against two real, independent competition results not seen during training: the author's own 645 kg total (90.6 kg bodyweight, 34, HPF) and a female competitor's 342.5 kg total (65.3 kg bodyweight, 26, HPF) — both fell within their respective predicted ranges. A pooled check across three recent Greek HPF meets in the same weight class (21 results) found only 2 exceeded the top of the predicted range, both well above the rest of the field.
Full plain-language walkthrough, including both validation examples, in
ATHLETES.md.
The tool in action — a real run, unedited, for a 90.6 kg, 34-year-old male HPF competitor:
===================================================================
Strength Standard Tool; Greek Raw PL
===================================================================
Athlete: Male, 90.6 kg, 34, HPF
===================================================================
Individual lift predictions
Squat 204.4 kg
Bench 135.0 kg
Deadlift 231.1 kg
===================================================================
Total predictions
Total 1; direct computation of total: 568.9 ± 162.4 kg
Total 2; computed as the sum of the individual lifts: 570.5 kg
Difference in the above estimates: 1.6 kg
→ The two models are consistent with one another
===================================================================
568.9 kg is likely between 406.5 and 731.2 kg; this range was checked
against real results, right about 95% of the time.
Fetch the data:
cd data/
curl -fL -o openpowerlifting-latest.zip \
"https://openpowerlifting.gitlab.io/opl-csv/files/openpowerlifting-latest.zip"
unzip -o openpowerlifting-latest.zip
mv openpowerlifting-*/openpowerlifting-*.csv .The notebook auto-detects the newest dated CSV in data/ — no code change
needed after fetching. To reproduce the exact results in this notebook, use
the 2026-08-01 release (hash 55149139, 778 MB).
Dependencies:
numpy
pandas
matplotlib
seaborn
scipy
scikit-learn
xgboost
optuna
shap
Install via:
pip install numpy pandas matplotlib seaborn scipy scikit-learn xgboost optuna shapRun: open greek_powerlifting_analysis.ipynb in JupyterLab and run all
cells from top to bottom. Cells have execution dependencies — do not run
individual cells out of order.
Try the tool: the final code cell of Phase 8 is self-contained — edit
the four # USER INPUTS lines at the top (bodyweight, sex, age, federation)
and run the cell to get a prediction. Do not modify anything below the
marked line.
| Model | Cohort | Features | Test RMSE |
|---|---|---|---|
| Predict-mean baseline | Total | — | 145.9 kg |
| Per-sex mean baseline | Total | — | 109.1 kg |
| Per-sex BW regression baseline | Total | — | 94.4 kg |
| OLS (normal equation) | Total | Base | 93.2 kg |
| OLS (normal equation) | Total | + Age | 87.1 kg |
| Ridge (optimal λ) | Total | Base | 93.2 kg |
| Ridge (optimal λ) | Total | + Age | 87.1 kg |
| Lasso (optimal λ) | Total | Base | 93.2 kg |
| Lasso (optimal λ = 1.79) | Total | + Age | 87.0 kg |
| XGBoost, untuned | Total | Base | 91.5 kg |
| XGBoost, untuned | Total | + Age | 82.2 kg |
| XGBoost, Optuna-tuned | Total | Base | 89.1 kg |
| XGBoost, Optuna-tuned | Total | + Age | 81.4 kg |
Uncertainty quantification (TotalKg, age-augmented, 95% nominal):
| Method | Interval half-width | Overall coverage | Male | Female |
|---|---|---|---|---|
| BayesianRidge (on Ridge) | ± 174.9 kg | 95.4% | 95.2% | 96.0% |
| Split-conformal (on tuned XGBoost) | ± 162.4 kg | 94.5% | 93.1% | 100.0% |
The male/female coverage asymmetry above is systematic — present at every confidence level, across all eight target/variant combinations, and under both uncertainty methods (see Phase 7). It reflects the underlying sex-scale heterogeneity in the data, not a flaw specific to either method.
All scratch implementations validated against sklearn to four decimal places. Per-lift (squat / bench / deadlift) results, subgroup breakdowns, and diagnostic plots — including SHAP age-dependence, Bayesian calibration curves, and conformal coverage tables — available in the notebook.
All competition data sourced from the OpenPowerlifting project (https://openpowerlifting.gitlab.io) under CC BY 4.0. Analysis and code by Kostas Rigatos, 2026.