Título del libro: Gecco 2022 - Proceedings Of The 2022 Genetic And Evolutionary Computation Conference
Título del capítulo: Multi-objective framework for quantile forecasting in financial time series using transformers
CARLOS IGNACIO HERNANDEZ CASTELLANOS; KATYA RODRIGUEZ VAZQUEZ;
Autores externos:
Idioma:
Año de publicación:
2022
Palabras clave:
Decision making; Deep learning; Electronic trading; Evolutionary algorithms; Financial markets; Forecasting; Neural networks; Time series; Decision makers; Decisions makings; Financial time series; Financial time series forecasting; Interval prediction; Multi objective; Multi-objectives optimization; Neural-networks; Time series forecasting; Uncertainty; Multiobjective optimization
Resumen:
The uncertainty associated with predictions is vital for decision making in financial time series forecasting. For the sake of enriching forecasts, quantile interval predictions are generated and their quality is evaluated using two conflicting indicators: quantile coverage error and quantile estimation error, which are later optimized using multi-objective evolutionary algorithms (MOEAs). The high performance quantile multi-horizon predictions are computed by an attention-based deep learning model based on transformers and the weights of its last layer are optimized using NSGA-II and NSGA-III. The Pareto fronts obtained in this work show that evolutionary algorithms can find a wide range of solutions that allow the decision maker to efficiently fine-tune the quantile forecasts without need of retraining the neural network. The results show that the decision maker can choose solutions whose risks' ranges variation is as high as 169% with only an increase of 5% in the original loss function. The dataset used in this work consists of S&P 500 Futures from Jan-2015 to Jun-2021 with one-hour frequency. © 2022 ACM.
Entidades citadas de la UNAM:
ISBN:
9781450392372
Editorial:
Association for Computing Machinery, Inc
