The economic time series prediction literature has seen an increase in research leveraging artificial neural networks (ANNs), particularly the multilayer perceptron (MLP) and, more recently, transformer networks. These ANN models have shown superior accuracy compared to traditional techniques such as autoregressive integrated moving average (ARIMA) models. The most recent models in the prediction of this type of neural network, such as recurrent or Transformers models, are composed of complex architectures that require sufficient processing capacity to address the problems, while MLP is based on densely connected layers and supervised learning. A deep understanding of the limitations is necessary to appropriately choose the ideal model for each of the prediction tasks. In this article, we show how a simple architecture such as the MLP allows a better adjustment than other models, including a shorter prediction time. This research is based on the premise that the use of the most recent models will not always allow better results.

Deep learning techniques have significantly advanced time series prediction by effectively modeling temporal dependencies, particularly for datasets with numerous observations. Although larger datasets are generally associated with improved accuracy, the results of this study demonstrate that this assumption does not always hold. By progressively increasing the amount of training data in a controlled experimental setup, the best predictive metrics were achieved in intermediate iterations, with variations of up to 66% in RMSE and 44% in MAPE across different models and datasets. The findings challenge the notion that more data necessarily leads to better generalization, showing that additional observations can sometimes result in diminishing returns or even degradation of predictive metrics. These results emphasize the importance of strategically balancing dataset size and model optimization to achieve robust and efficient performance. Such insights offer valuable guidance for time series forecasting, especially in contexts where computational efficiency and predictive accuracy must be optimized.

In this paper, we analyse two interesting applications related to the dynamics of economic phenomena linked to the Efficient Market Hypothesis (EMH), informative surprises, and the Model-Data Paradox of Chaos in certain top currency pairs from the foreign exchange market (FOREX). On the one hand, we empirically show that the FOREX market reacts under the Efficient Market Hypothesis in some cases, creating a significant variation in a short period of time (15, 30, and 60 min) in the quotes of the main currencies from the most important economic regions in the West (the United States, Europe, and the United Kingdom). This variation would depend on the actual deviation of high-impact macroeconomic news reported by these markets in relation to trade balance, unemployment rate, Gross Domestic Product (GDP), retail sales, the Industrial Production Index (IPI), and the Consumer Price Index (CPI). On the other hand, by testing the Model-Data Paradox of Chaos, we empirically verify that if we consider all the information available in the financial markets of currencies (or at least, more desegregated data) instead of daily data, and we apply a robust chaotic behaviour detection method, we can find differences in relation to the detection of chaos on the same series but with different temporal frequencies. This allows us to confirm that behind these financial time series which show an apparently random irregular evolution, there would be a generating system which, although unknown in principle, would be deterministic (and nonlinear), and we could take advantage of that deterministic character to make predictions, even if only in the short term, understanding “short term” as the time it takes for the market to incorporate these informative surprises in the FOREX market analysed.