Predicting Chinese lottery numbers with sequence models

What it does This project scrapes historical winning numbers for China’s 双色球 and 大乐透 lotteries from 500.com, then trains separate deep-learning models to predict upcoming draws. Red balls are handled as a sequence prediction task, while blue balls are modeled separately, with results available via command-line scripts or a web interface. Model hyperparameters and train-test splits live in config.py.

The interesting bit The author ditched independent-ball assumptions in favor of a sequence model with a CRF layer, mainly to eliminate duplicate red-ball predictions. The catch: the CRF implementation required disabling TensorFlow 2 eager execution and falling back to tf.compat.v1 APIs, a compromise the README notes openly.

Key highlights

• Supports both 双色球 (ssq) and 大乐透 (dlt) via a runtime flag
• Uses a sequence model with CRF layer to avoid duplicate red-ball predictions
• Web UI at ai-spinach.xyz for training and prediction without local setup
• Auto-scrapes historical draw data from the 500.com lottery site
• Configurable model hyperparameters and train/test split ratios in config.py

Caveats

• Data ingestion depends on scraping 500.com, so site changes or blocks break the pipeline
• TensorFlow runs in 1.x compatibility mode inside TF 2.x, which is technical debt
• The README documents evaluation features but cites no accuracy metrics or performance baselines

Verdict Developers curious about applying CRFs to short tabular sequences, or Chinese lottery players who want an algorithmic number picker, might enjoy tinkering. Anyone seeking a proven statistical advantage—or who does not play these specific lotteries—can safely skip it.