Complexity-Based Lambda Layer for Time Series Prediction

Time series analysis forms the basis for the temporal sequences that we observe in everyday natural phenomenon. Examining and characterizing time series' forms the basis for research in areas spanning classifying heart rate variability, temperature prediction and stock price prediction. In recent history combining powerful techniques such as Fourier Analysis with Machine Learning techniques has improved our ability to predict future time series based on historic data. Complexity is one of such tools that incorporates long-range dependency based on the inherent self-similarity that exists in many natural phenomena. Leveraging the prowess of recurrent neural networks with that of complexity measures combines two very powerful techniques to improve prediction accuracy. In this work a Complexity Lambda Layer was initialized in series with Artificial Neural Networks (ANN) architectures to improve prediction accuracy for synthesized Brownian Noise. Window size for recurrence and the stationary interval size was optimized for increased performance. For both non-temporal and temporal studies, a 2-4 fold improvement in root-mean-squared accuracy was obtained. This approach was implemented as a Lambda layer, meaning the improvement can be done on-the fly with minimal overhead.