Dynamic Neural Networks for Daily Allocation Across Bank Stocks

Summary

The project proposes a daily allocation strategy across stocks in India’s Nifty Bank index. It uses three dynamic linear neural network models: one estimates the probability of a positive next-day return, while two estimate return size conditional on positive or negative outcomes. Inputs include recent stock and index returns, price and volume changes, volatility, moving-average and range measures, technical indicators, and stock-index correlation. The models are updated each trading day, and log returns are used as targets to support stationarity.

Expected returns are formed from the probability and conditional return estimates, then scaled by recent volatility to rank holdings. Cash is allocated among stocks with positive scores; positions with nonpositive scores are reduced or sold, and a 20% stop-loss is applied per stock. The article reports a backtest that later outperformed its benchmark and a short-period return, but acknowledges that transaction costs are excluded and execution quality matters. The data and model details shown are incomplete, so these reported results do not establish robustness or live performance.

Key ideas

• Three dynamic linear neural networks estimate return direction and conditional positive and negative return sizes.
• Expected returns are combined with recent volatility to determine relative stock allocations.
• The strategy reallocates cash daily among positively ranked bank stocks and reduces negatively ranked positions.
• A per-stock stop-loss is used, while transaction costs and execution effects are absent from the backtest.
• The reported performance is limited evidence and does not demonstrate durable live profitability.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.