Energy-Efficient Neuromorphic Spiking Neural Networksfor Real-Time Event-Driven Stock Market Prediction: A Systematic Review

Abstract

I believe that neuromorphic computing and financial prediction is now a genuinely exciting, yet under-researched, area of applied machine learning study. The paper presents a systematic literature review of the literature based on energy efficient Spiking Neural Networks (SNNs) in forecasting events-based real-time stock market data. Event based in character Financial time-series data is not continuous in nature, but consists of major discrete events like significant price changes, volume surges, indicator shifts, etc. This is a structural factor that makes SNNs, the only neuron model that grows the input events, a theoretically optimum architecture to process financial signals, the Leaky Integrateandfire (LIF) neuron model. We survey work in the fields of deep learning to financial forecasting, SNN training algorithms (including STDP, surrogate gradient methods and ANN-to-SNN conversion), neuromorphic hardware problems, and the small but growing number of literature applying SNNs to financial data. Significant works in IIT Kharagpur, IIT Bombay and IIT Delhi are discussed too as well as the standard foreign works. The review identifies four gaps in research, including standardised SNN benchmarks on financial data, the analysis of the real implementation of hardware to financial SNNs, the analysis of multimodal SNN architecture to financial inputs, and the near-complete absence of online learning protocols which are suitable on non-stationary market environments. These areas of gaps are visible and concrete and in which future research can be done on the frontier of neuromorphic financial intelligence.