CXu0630/mtg-price-catboost

MTG card price range: CatBoost quantile baseline

Magic: The Gathering is a trademark of Wizards of the Coast, LLC. This is an unofficial fan project, not produced or endorsed by Wizards of the Coast.

Source code, training pipeline and design doc: mtg_price_predict_catboost_deep_gp

Predicts an 80% price range (USD) for a Magic: The Gathering card printing (one card in one set/treatment/finish) from 93 tabular features: rarity, finish, frame/border treatment, mana cost, type, P/T, legalities, booster pull rates, EDHREC rank and artist. This is the project's "Model 1", a fast CPU-only baseline. The deep-kernel GP is CXu0630/mtg-price-deep-gp.

Method

• CatBoost MultiQuantile regression (p10 / p50 / p90) on log(price_avg). Settings: depth 6, l2_leaf_reg 3, learning rate 0.08, early-stopped at iteration 2,709.
• "Tier A" (fundamentals-only) feature set. In an ablation, adding ~350 sparse community tag columns made coverage worse in every price bucket.
• The training split was rebalanced by capping each of 30 log-price bins at 8,000 rows, so bulk commons don't dominate the loss. Calibration and test splits keep the natural price distribution.
• Group-conditional split-conformal calibration fits a separate additive correction to the [p10, p90] interval for each predicted-median bucket (<$1, $1–10, $10–100, $100+) on a held-out calibration split.
• set / set_name are deliberately not features, so the model generalizes to sets it hasn't seen.

Results

Held-out test split (28,377 printings, 20%), split by oracle_id so that no card appears in both train and test. Log-price MAE of the median; coverage is the calibrated 80% interval.

true price	n	MAE (log)	80% coverage	mean width (log)
< $1	17,179	0.42	77.6%	1.40
$1–$10	7,641	0.61	82.9%	2.08
$10–$100	3,236	0.74	71.8%	2.05
$100+	321	1.14	52.3%	2.03
overall	28,377	0.52	78.1%	1.66

Compared with the deep GP, this model has lower median error, but its intervals under-cover expensive cards ($10–$100 and $100+). Each model was evaluated on its own split, so the comparison is approximate.

Usage

# pip install catboost huggingface_hub pandas pyarrow
import sys
from huggingface_hub import snapshot_download

path = snapshot_download("CXu0630/mtg-price-catboost")
sys.path.insert(0, path)  # the repo bundles its inference code in mtg_price_prediction/
from mtg_price_prediction.catboost_model import CatBoostPredictor

model = CatBoostPredictor.from_pretrained(path)
ranges = model.predict(features_df)  # -> price_low / price_mid / price_high (USD)

Or with plain CatBoost:

import json, numpy as np
from catboost import CatBoostRegressor
from huggingface_hub import snapshot_download

path = snapshot_download("CXu0630/mtg-price-catboost")
cfg = json.load(open(f"{path}/config.json"))
model = CatBoostRegressor(); model.load_model(f"{path}/model.cbm")

X = features_df[cfg["feature_cols"]].copy()
for c in cfg["cat_features"]:
    X[c] = X[c].fillna("__missing__").astype(str)
p10, p50, p90 = model.predict(X).T
buckets = np.digitize(p50, np.log([1, 10, 100]))   # 0:<$1 1:$1-10 2:$10-100 3:$100+
corr = np.array([cfg["calibration_corrections"][str(b)] for b in buckets])
low, mid, high = np.exp(p10 - corr), np.exp(p50), np.exp(p90 + corr)

features_df has rows shaped like the project's prepare_training_data.py output. config.json → feature_cols lists the exact 93 columns.

Files

file	contents
model.cbm	CatBoost model (native format)
config.json	feature columns, categorical columns, hyperparameters, calibration corrections, test metrics
mtg_price_prediction/	inference code (CatBoostPredictor, calibration)

Limitations

• Expensive cards are under-covered. Calibrated coverage is 52% at $100+ and 72% at $10–$100, against an 80% target. The calibration sample for $100+ is small (131 points).
• The model goes stale. It is a direct regressor on absolute log-price, trained on a ~3-month TCGplayer window ending September 2026, so it drifts with the market.
• Results are from a single seed and a single split.
• Out of scope: multi-face cards, digital-only printings, and oversized cards.

Training data

CXu0630/mtg-card-market-dataset has one row per (printing, finish), 142,012 of them with a price. It is built on pcwoods/mtg-card-prices (Scryfall card data + TCGplayer price history) and joined with booster pull rates from CXu0630/mtg-print-distribution.