Phillip K.S. Chu
Jul 22, 2017 · 2 min read

Packing multiple data streams into n-dimensional matrix is often a necessary step in data preprocessing, before feeding into any neural network, recurrent or convolutional. But it doesn’t automatically represent learnable spatial patterns, that CNN is suitable for.

Using OHLCV to predict Up and Down in a time series, that is the “nail”, a nail with temporal locality. Why is CNN the better hammer here?

This simplistic repackaging of OHLCV → Up/Down into some n-dimensional matrix, when super-imposed into an image, where is the temporal relationship, to be learned by a CNN?

Your reason for using CNN is “because of flexibility and interpretability of hyperparameters (convolutional kernel, downsampling size etc) and performance similar to RNNs, better than MLP with much faster training.” — — not because a strong belief that the OHLCV time-series, when trivially repacked as such, becomes a more learnable candidate for CNN, than for RNN. Or am I missing something there?

There are discussions on applying CNN to time series, specifically a pretty good paper on Encoding Time Series as Images for Visual Inspection and Classification Using Tiled Convolutional Neural Networks

Although your CNN yields “improvement in accuracy” over the previous RNN set up, have you tried exploring LSTM, architecture and hyperparameters, to handle the multivariate OHLCV? RNN is a more natural candidate in general for time series problems, uni- or multivariate. In fact, the input n-dimensional matrix is identical for an LSTM.

Philippe Remy has an in-depth discussion on configuration possibility of LSTM for time series forecasting.

It would be good to know where the difference in accuracy come from. Is it because a poorly tuned RNN, or some ground-breaking truth that “time series is better predicted by CNN!!”, that the whole financial industry should heed (in which case, congratulations, on discovering the best-kept secret in the financial market!)

)

    Phillip K.S. Chu

    Written by

    Urban Nomad・Street Photog・Less is More・Antifragility・Stoicism