MultiAsset Market Making Model: Transformer based Interval Forecasting
Model Summary & Evaluation
Overview
This project implements a transformer based time-series forecasting model for 5-minute OHLCV data on Binance BTC Perp and OKX BTC Spot. The model uses cross-channel attention and a custom interval score loss to predict high/low ranges for both assets.
Model achieves 47.35% (BTC Perp) and 74.43% (BTC Spot) interval coverage accuracy with lambda_width=5.
Prediction vs True Value (Visualization)
Prediction vs True Value
Technical Foundation
Python (pandas, matplotlib, mplfinance): Data download, cleaning, plotting
PyTorch: Model architecture, training, evaluation
MultiheadAttention (5 heads): Cross-channel attention for multi asset modeling
Interval Score Loss (lambda_width=5): Penalizes missed intervals and wide ranges, differentiable for training
Data alignment: Merging Binance and OKX OHLCV by timestamp
Codebase Status
main.py: Trains model on combined Binance/OKX data, interval score loss, outputs 4 values: BTC Perp high/low and BTC Spot high/low
test_only.py: Plots true/pred candlesticks for both exchanges, computes relaxed (OR) interval accuracy
Problem Resolution
Issues: Data ordering, deduplication, plotting errors, shape mismatches, loss function encouraging wide intervals
Solutions: Batch reversal, sorting, deduplication, robust CSV reading, interval score loss with width penalty, OR logic for interval accuracy
Lessons Learned: Interval score loss must balance coverage and precision; width penalty is critical to avoid trivial solutions
Progress Tracking
Completed: Data download/cleaning, plotting, model implementation, training, evaluation, documentation
Recent Context: Summary and evaluation based on latest training/test results and loss function
Pending: Further model tuning (lambda_width, architecture), additional evaluation metrics or visualizations, publishing code/model/dataset
Disclaimer
Note: This repository is a work in progress. The model weights, code, and dataset will be published in a later stage. Current release includes only documentation, summary, and evaluation logs. Stay tuned for updates!
