G-Research Crypto Forecasting

Patrick Yam

Final leaderboard

Competition setup

• 2 stages competition
• Training phase: 2nd Nov 2021 – 1st Feb 2022
• Evaluation phase: 2nd Feb 2022 – 3rd May 2022
• The model is fixed after the training phase and will be run against real market data in the evaluation phase

Overview

• Data: 14 crypto assets from Jan 2018 to Jan 2022, minute bar data

Minute bar data

Target

Evaluation metric

• Weighted version of the Pearson correlation coefficient

​

​

Data preprocessing

• Additional features: sin/cos time of the day
• Model inputs: 9 features in total
• ['Count',  'Open',  'High',  'Low',  'Close', 'Volume', 'VWAP', 'time_sin', 'time_cos']

​

• Take latest 45/60/90/120 minutes as input

​

• Log transform for 'Volume' and 'VWAP'
• Perform standardisation locally (except for time_sin and time_cos)

Model Architecture

Axial Attention

• Ordering matters for time series axis but not for asset axis

For each asset, get information from other timestamp

For each timestamp, get information from other asset

MLP pooling

• Nonlinear version of weighted average pooling
• Perform MLP for time series axis, with output dimension = 1

​

• Layer:

​

• Forward:

​

​

​

Loss function

• Minibatch Negative Weighted Pearson correlation coefficient

​

• The scale for model's prediction trained with different random seed could be very different 
• [0.1, 0.2, 0.3] will give the same result as [10, 20, 30]
• Could be problematic for ensemble

​

• BatchNorm with 'affine=False' for model's prediction

Ensemble

Correlation between models trained with different sequence length

• 3-fold CV training
• 4 different sequence length settings
• Total 12 models, simple average

​

​

• Single model performance:�Private LB: 0.0202 -> 0.0159�#28, silver medal

Longer sequence doesn't always come with a better result

Thank you!��Q&A