Short-Term Volatility Prediction Using Deep CNNs Trained on Order Flow
Research report
TLDR
Order flow data is converted to 2D images and fed into convolutional neural network (CNN). CNNs generate 1-minute crypto volatility forecasts which are compared to canonical forecasting models.
Key learnings
This paper adds to the growing corpus of academic literature utilizing CNNs in financial forecasting scenarios. While still mostly relegated to academic studies, CNNs are increasingly being used in practice for non-image processing applications.
Of particular interest to Concrete is the methodology used to discretize time series feature into 2-dimensional image representations. This approach is applicable to a range of features which and allows for efficient encoding of information with low signal-to-noise ratio.
Methods and outputs
High-level schema of the feature encoding algorithm, above (page 5). The data is encoded via pixels – each image has multiple channels encoding specific information between time and value intervals.
Encoding column values generally represent time for the other variable, which is converted to a scale from 0 to 255. The horizontal values encode discrete channels which are orthogonal to one another.
Example of the encoding. The pixel values in the images represent the normalized volume of limit orders within the corresponding interval – the intensity of each pixel represents the level of order volume at a particular time and price level.
Testing results show the pixel-based encoding and CNN approach works well as compared to other forecasting models (GARCH, LSTMs, etc.) according to Root Mean Square Percentage Error (RMSPE) values.
Concrete
As Concrete’s risk models expand to incorporate increasingly long-tail features, CNNs offer an attractive way to encode and compress information.
CNNs offer an easy way to initially test features for orthogonality and correlation in a visually digestible manner. This could be useful to help non-technical team members test hypotheses.
Links and further reading
2020 paper utilizing CNNs for US equities trading. This paper has not only academic results, but some compelling results in practice.
Last updated