Predicting Price Movements in High-Frequency Financial Data with Spiking Neural Networks

Battle Count: 71

Highly relevant to quantitative trading. The paper directly addresses high-frequency price spike prediction, a core challenge in HFT. It provides a complete pipeline from data preprocessing through model training to backtested trading strategy evaluation. The PSA metric and Bayesian optimization framework are practical contributions. The unsupervised STDP approach offers potential advantages for real-time deployment without labeled data. However, the single-stock, short-period evaluation and lack of transaction cost modeling limit immediate practical applicability.

High complexity due to: (1) SNN simulation with LIF neuron dynamics requiring custom implementation, (2) STDP learning rule with exponential windows and synaptic homeostasis, (3) Poisson spike encoding pipeline, (4) Bayesian optimization over high-dimensional hyperparameter spaces (13+ parameters), (5) Multiple model architectures with different training paradigms (unsupervised STDP vs supervised BPTT), (6) Novel PSA objective function, (7) Rolling window backtesting framework. Requires expertise in both computational neuroscience and quantitative finance.

The paper claims a 'fully reproducible, end-to-end training framework' and provides detailed algorithm pseudocode, hyperparameter search spaces, and preprocessing steps. However, no GitHub repository or code link is provided. The use of publicly available AAPL tick data from Feb 2015 aids reproducibility, but the specific data source and exact preprocessing implementation details would need to be inferred.