MTG card price range: deep kernel GP

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_prediction

Predicts an 80% price range (USD) for a Magic: The Gathering card printing (one card in one set/treatment/finish) from what's printed on the card. It works for released cards and for hypothetical or unreleased ones. It is the second opinion to the project's boosted-tree fair-value model (Model A): it sees the same card attributes and supply inputs, and also reads the card's rules text and art.

Architecture

• Wide & deep feature extractor (PyTorch). Each input is projected separately and the results are concatenated into a 48-dim joint embedding:
  • Frozen oracle-text embedding: sentence-transformers/all-mpnet-base-v2, 768 → 64.
  • Frozen card-art embedding: CLIP ViT-B/32, ViT-B-32-quickgelu / openai weights in open_clip, run on Scryfall's art_crop; 512 → 64.
  • One nn.Embedding per categorical column (rarity, finish, frame, artist, ...).
  • An MLP over Model A's boolean/numeric inputs: mechanics, keywords, top-N oracle/art/type tags, legalities, EDHREC rank, booster pull rates and supply (set size, rarity-slot size, products carrying the printing), mana-cost detail, reprint and age history, and the set's price level (mean and top-5 mean log price of the set's other training cards).
• Sparse variational GP (GPyTorch SVGP) with 750 inducing points and an RBF-ARD kernel over the joint embedding, Gaussian likelihood. Trained end to end on the ELBO against log(price_avg).
• Group-conditional split-conformal calibration. The raw mean ± 1.2816·std interval gets a separate additive correction for each predicted-price bucket (<$1, $1–10, $10–100, $100+), fit on a held-out calibration split.
• Training used graduated oversampling of the $100–$1000 range to improve coverage on expensive cards.

Results

Held-out test split (14,639 printings), split by oracle_id so that no card appears in both train and test. Log-price MAE; coverage is the calibrated 80% interval.

Compared with the boosted-tree fair-value model (Model A: log MAE 0.43, 80.0% coverage overall, 65.4% at $100+), this model has the wider midpoint error but better coverage on expensive cards. Its gain comes from its calibrated intervals, not sharper point estimates. Each model was evaluated on its own held-out split, so the comparison is approximate. Exact numbers are in config.json → test_metrics.

Permutation importance on the test split (rise in log MAE when a group is shuffled): EDHREC rank +95%, pull rates +46%, card/printing age +33%, finish +26%, set price level +23% (+45% on $10+ cards). The text and art embeddings add +4% and +3%.

Usage

# pip install torch gpytorch safetensors huggingface_hub pandas pyarrow
import sys
from huggingface_hub import snapshot_download

path = snapshot_download("CXu0630/mtg-price-deep-gp")
sys.path.insert(0, path)  # the repo bundles its inference code in mtg_price_prediction/
from mtg_price_prediction.deep_gp import DeepGPPredictor

model = DeepGPPredictor.from_pretrained(path)
ranges = model.predict(features_df, text_emb, image_emb)  # -> price_low / price_mid / price_high (USD)

To build features_df from rows of CXu0630/mtg-card-market-dataset:

from mtg_price_prediction.gp_features import gp_feature_frame
features_df = gp_feature_frame(rows, model.prep["wide_cols"], model.feature_context(dataset))

feature_context supplies what the Model A inputs need: the full dataset (set sizes, each card's other printings), set_prices.parquet, and the as-of date for age features (config.json → model_a_inputs).

• features_df has rows shaped like the project's prepare_training_data_model2.py output: the columns listed in preprocessor.json (categorical_cols + wide_cols). Missing numeric values are imputed. Unseen categories map to __unk__.
• text_emb is (n, 768), the card's oracle text (plus any token or meld-result text it creates) run through all-mpnet-base-v2.
• image_emb is (n, 512), the CLIP ViT-B-32-quickgelu image embedding of the art crop.
• pull_rate_defaults.parquet gives the expected booster pull rate for each (rarity, finish). Use it to fill *_rate_percent for a card that isn't in a real product yet.

Files

Limitations

• $100+ is under-covered. Calibrated coverage there is 71.4%, not 80%, and the bucket is small (182 test points). Prices driven by combos, bans or collectibility are mostly invisible to these features.
• Hypothetical cards lose the set price input. A card from a set with no priced training cards gets set_price_missing and an imputed level.
• The GP's uncertainty is nearly constant across cards. Predictive std varies by only about 1% within a bucket, a known deep-kernel-learning collapse. Interval width adapts by predicted price bucket, not per card.
• Most of the signal is tabular. A tabular-only variant comes close to this model, an embeddings-only one collapses, and the embeddings add only a few percent in permutation importance.
• Results are from a single seed and a single split. Differences of a few points between variants are within noise.
• The model goes stale. It was trained on a ~3-month TCGplayer price window ending September 2026 and predicts absolute log-price, so it drifts as the market moves.
• Out of scope: multi-face cards (transform, MDFC, split, adventure, ...), 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. Card art comes from Scryfall.