UFC-Almanac

A collection of datasets of UFC fight results, fight stats (both updated weekly) and fighter data (updated monthly), containing the data for all UFC fights since 2010. Also contains pipelines for transforming the data into formats for training machine learning models, and training scripts for a variety of deep learning models.

Next UFC Event Predictions

Event date: October 3, 2026

Fight	Win				Loss				Draw
	Overall	KO	Sub	Dec	Overall	KO	Sub	Dec
Natalia Silva vs Wang Cong	58.8%	9.8%	6.1%	42.9%	40.7%	7.7%	3.1%	29.9%	0.3%
Deiveson Figueiredo vs Payton Talbott	33.4%	6.4%	7.4%	19.6%	66.1%	27.5%	7.0%	31.6%	0.6%
King Green vs Esteban Ribovics	39.6%	7.8%	5.6%	26.2%	59.9%	25.7%	3.7%	30.5%	0.6%
Ateba Gautier vs Roman Kopylov	67.4%	31.2%	7.4%	28.8%	32.2%	15.0%	4.0%	13.2%	0.6%
Imanol Rodriguez vs Alden Coria	47.9%	7.4%	5.8%	34.7%	51.8%	7.5%	6.0%	38.3%	0.4%
Damian Pinas vs Andrey Pulyaev	49.9%	23.1%	11.1%	15.7%	49.6%	21.9%	8.7%	19.0%	0.6%
Marcus McGhee vs Benardo Sopaj	37.8%	10.6%	4.0%	23.2%	61.8%	12.2%	10.1%	39.5%	0.4%
Johnny Walker vs Mick Parkin	30.8%	16.0%	4.3%	10.5%	68.7%	38.3%	8.1%	22.3%	0.5%
Rafael Dos Anjos vs Alexander Hernandez	44.9%	4.9%	7.5%	32.5%	54.7%	11.9%	3.1%	39.7%	0.4%
Marvin Vettori vs Ismail Naurdiev	38.2%	4.7%	5.0%	28.5%	61.2%	4.8%	4.3%	52.1%	0.5%
Court McGee vs Eric Nolan	51.8%	3.1%	28.0%	20.7%	47.5%	19.0%	1.7%	26.8%	0.6%

The model used for these predictions is a models/transformer_model.py trained using the following command:

ufc-train --model transformer --path artifacts/core/transformer_model.pt --epochs 75 --dropout 0.5 --num-layers 2 --restarts 10 --d-model 32 --learning-rate 1e-4 --weight-decay 1e-4

Disclaimer: These predictions are generated by a machine learning model and reflect estimated probabilities, not certainties. They are provided for informational and entertainment purposes only and should not be used as betting or financial advice.

Setup

pip install -e ".[all]"

To scrape data, install the Playwright browser:

playwright install chromium

Usage

After installing the project, you can run commands either via the console scripts or the scripts/ entrypoints.

Scrape data (into data/*.csv files)

# Scrape fight results and stats
ufc-scrape-fights
# or
python scripts/scrape_fights.py

# Scrape fighter profiles
ufc-scrape-fighters
# or
python scripts/scrape_fighters.py

Training

ufc-train --model transformer --rebuild-data
# or
python scripts/train.py --model transformer

Parameters

Flag	Description	Default
--model	Model architecture (linear, mlp, or transformer)	linear
--epochs	Number of training epochs	40
--batch-size	Training batch size	256
--learning-rate	Adam learning rate	3e-5
--val-fraction	Fraction of the most recent samples held out for validation	0.1
--weight-decay	L2 regularization strength for Adam	3e-5
--dropout	Dropout probability	0.5
--d-model	Transformer hidden dimension (transformer only)	128
--num-layers	Number of transformer encoder layers (transformer only)	4
--max-fights	Past fights per fighter / sequence length (transformer only)	8
--path	Path to save trained model weights	<ModelName>.pt
--rebuild-data	Regenerate training data from CSV files	off
--optimize-temp	Optimize temp scaling on the val set	off
--restarts	Number of independent training runs	1
--brier-weighting	Weight on val Brier vs loss when selecting checkpoints	2

Use --rebuild-data when the underlying CSV data has been updated. Changing --max-fights also regenerates transformer training data when it does not match the saved tensors.

Inference (predict fight outcomes)

Interactive CLI:

ufc-predict --model linear
# or
python scripts/predict.py --model mlp

Enter two fighter names when prompted. Type exit, quit, or q to stop.

Parameters

Flag	Description	Default
--model	Model architecture to load: linear, mlp, or transformer	linear
--path	Path to trained model weights	<ModelName>.pt

The predictor loads trained weights and normalization stats from artifacts/checkpoints/ for the selected model by default. Train a model first with ufc-train, or pass --path to load a custom checkpoint.

Automated data updates

GitHub Actions workflows scrape new fight data weekly and fighter data monthly, adding new data to the csv files in the data/ directory.