Ethereum’s Alphabet : Predictive Price Chart Patterns Discovered by a Neural Network

Based upon the recent work of Dacheng Xiu's team at the University of Chicago, we demonstrate that we can repurpose a consumer-level character recognition neural network to recognize price chart patterns in the Ethereum cryptocurrency for the purpose of forecasting short term (small N days) bullish/bearish signals (binary) for the same currency. Chart patterns are machine-interpreted only; any prior knowledge or biases about chart patterns are not applied in this study. Out of sample forecasts are found to be notably greater than 50% accurate and statistically significant for a limited range of days forward. We then begin to explore multi-asset candidate predictors of Ethereum returns (in particular, Bitcoin and Ripple cryptocurrency price chart patterns) using the same image recognition approach, and we start to check model quality as a function of time. Due to time-dependent degradation of the model, we rebuild a simpler model using the same data but arranged to allow linear-in-the-coefficients binary logistic regression models to be fit. These models also degrade with respect to out-of-sample forecasts when time-shifted, but good results in some time windows suggest that simpler linear models should be explored in more detail versus the image recognizers. Further, we discovered that allowing non-linearities such as fractional exponents and time shifts into a standard logistic model improved historical time-shifted results out of sample