NON-LINEAR MODELLING IN BRAZILIAN MARKET: EVALUATING THE FORECASTING PERFORMANCE OF NN (UNIVARIATE NEAREST NEIGHBOR) AND SNN (SIMULTANEOUS NEAREST NEIGHBOR) FORECASTING ALGORITHM

• Autor(a) principal: Scherer Perlin, Marcelo
• Data de Publicação: 2013
• Outros Autores: Ceretta, Paulo Sérgio
• Tipo de documento: Artigo
• Idioma: por
• Título da fonte: REAd (Porto Alegre. Online)
• Texto Completo: https://seer.ufrgs.br/index.php/read/article/view/39956
• topic: Non linear Forecasts; Univariate and Multivariate Nearest Neighbor; Market Efficiency

The predictability of stock market’s behavior is a topic studied by different academic circles for long time. A popular tool to make predictions about the stock market behavior on short term is the technical analysis. Such tool is based on the analysis of quantitative indicators and also chart patterns in order to identify the time to entry (buy) or exit the market (sell). A quantitative approach that is related to charting is the use of the non-parametric approach of nearest neighbor algorithm in order to produce forecasts of the time series on t+1. The main objective of this paper is to study the forecasting performance of the nearest neighbor method for the Brazilian Equity data in two versions, the univariate and also the multivariate case, which is also called simultaneous nearest neighbor. The main conclusion of the paper is that the ability of the algorithm in forecasting the values of the stock prices is mixed. A comparative analysis with the random walk model showed that this naïve approach has more explicative power in numerical accuracy. For the case of directional forecasts, the NN presented better results, resulting in correct directional forecasts moderately higher than 50% for most of the assets and with a maximum of approximately 60% correct market direction forecasts, which indicates that the method may add value in quantitative trading strategies. Comparing the results for both versions of the algorithm, its clear that both presented very similar results, but the univariate case was slightly better.