580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How do different neural network architectures compare in forecasting financial time series data?
This research area investigates the comparative performance of various neural network architectures, such as feedforward networks, recurrent neural networks (RNNs), long short-term memory (LSTM), gated recurrent units (GRUs), higher-order neural networks, and hybrids with traditional models in forecasting financial time series like stock prices, exchange rates, and market indices. It matters because understanding which architecture better captures nonlinear and temporal dependencies improves forecast accuracy in volatile financial markets.
Key finding: This study compared feedforward and recurrent neural networks trained using backpropagation and extended Kalman filter for exchange rate forecasting (EUR/RON and USD/RON). It found that recurrent architectures with advanced... Read more
Key finding: By evaluating RNN and its variants LSTM and GRU on stock index and currency exchange rate data from diverse economies, the paper showed the GRU architecture generally achieves superior out-of-sample forecasting accuracy... Read more
Key finding: Proposed a second-order neural network with GA-tuned parameters that achieved better forecasting accuracy than recurrent neural networks and multilayer perceptrons on one-day-ahead stock closing price prediction. This... Read more
Key finding: Using real-world stock market data, the study demonstrated that LSTM networks effectively model temporal dependencies and improve prediction accuracy of stock prices over traditional models by utilizing cross-validation and... Read more
Key finding: Compared ARIMA and feedforward neural networks for forecasting stock market capitalization in Dhaka Stock Exchange, concluding that a three-layer feedforward ANN with backpropagation outperformed ARIMA (2,1,2) in forecasting... Read more
2. In what ways can hybridizing neural networks with classical financial forecasting models improve prediction in volatile markets?
This research theme explores the integration of neural networks with traditional statistical models like ARIMA, exponential smoothing, GARCH, and variants thereof, to create hybrid models that incorporate both linear and nonlinear data aspects. The focus is on leveraging complementary strengths to improve forecasting accuracy, especially in volatile, nonlinear, and noisy financial time series.
Key finding: Developed a hybrid model combining exponential smoothing (a linear method) and artificial neural networks (capturing nonlinearities) via linear programming, showing superior prediction accuracy on real-world financial time... Read more
Key finding: Compared classical ANN models (MLP, RNN) with hybrid networks incorporating GARCH and EGARCH inputs, finding that hybrid GARCH-ANN and EGARCH-ANN models displayed improved prediction performance (lower MSE and MAD) on... Read more
Key finding: Included an ARIMA-PNN hybrid approach where ARIMA residuals' trends were modeled by probabilistic neural networks to adjust forecasts, illustrating improved performance over standalone ARIMA models. This highlights how... Read more
3. How can feature selection and variable importance techniques enhance neural network-based financial forecasting?
This theme addresses incorporating data mining and machine learning methods, such as information gain and variable reduction techniques, to select relevant predictors for ANN models in financial forecasting. By focusing on variable importance, the approach seeks to improve model generalization, reduce overfitting, and enhance forecasting accuracy and robustness.
Key finding: Employed an information-gain technique to evaluate over numerous financial and economic variables, reducing dimensionality before training neural networks for forecasting stock returns. The classification models trained on... Read more
Key finding: Found that including more variables beyond key predictors (trading volume and loan interest rates) did not improve stock return forecast accuracy. Incorporating a focused set of explanatory variables in both regression and... Read more
Key finding: Systematic literature review found that artificial neural networks, often combined with variable selection strategies like genetic programming and particle swarm optimization, outperform other models in exchange rate... Read more