Nearest-neighbor Predictions In Foreign Exchange Markets

In the age of economic globalization, realistic analysis and forecasting of exchangerate fluctuations is very important for policy formulation and investmentdecision-making. However, the volatility of the exchange rate is affected by manyuncertainties. It is difficult to find a precise fundamental model which is effectiveenough to describe exchange rate fluctuations and their links between the factors.Phenomenal forecasting models based on nonlinear statistical patterns in the exchangerate time series are a natural alternative pathway. In this paper we examine the featurespace of exchange rate short-term return series. We use phase-space reconstructiontechniques and semi-supervised affinity propagation clustering to establish nearestneighbor predictions. The model is tested using the recent10-year daily data ofEUR/USD. In the end, we use mean square error and hit rate to measure the predictionaccuracy. The test result shows that the model has exhibited high prediction accuracy,thus supporting the validity of the model. The main works are as follows:1．Through analyzing the characteristics of the foreign exchange market and therole of causal model in the modern finance, foreign exchange market, as an integral partof financial market, is a complex economic system which contains a large number ofuncertain factors. There is the mutual contact and mutual influence between the factorsin such a complex economic system. The interaction between the factors is result ofchaotic phenomenon in the foreign exchange market. Due to integrity of the historicaldata and convenient access to data in financial markets, value assessment and riskmeasure model is based on probability as the core concept of mathematical statistics,mainly includes: association, correlation and likelihood. Most financial economistsconsider that history will repeat someday, that is to say the future market behavior willrepeat the past happened in market behavior. So while available quantitative theoriesand models for finance prefer correlation, they most ignore the causality which willproduce unexpected results.2．Through analyzing the chaos theory and phase space reconstruction theory andcombined with the characteristics of the foreign exchange market, we apply them to the prediction of foreign exchange. Chaos is a likely random phenomenon which is acommon complex movement in deterministic system. Because the chaos phenomenon isbetween sure and stochastic, it invariably associates with predictions and we needobservational data to predict. The phase space reconstruction is the premise of chaosresearch and calculation which will be subject to limited low-dimensional data toreconstruct high dimensional dynamical systems.3．Through analyzing classification and clustering algorithm, the classificationprocess is analysis of the given data. We use the information obtained from the analysisof data for classification and describe the classification accurately. Then we classify thefuture possible data of database. It is a good method for clustering which mainly lies inthe definition of unknown classification schema, classification model including thenumber of classes and the class description. So clustering problem research becomeshow to determine the number of classes and the class description. Semi-supervisedaffinity propagation clustering introduces cluster validity index which evaluates thequality of the clustering result and supervise iteration process.4．According to characteristics of foreign exchange market, we establish anabstract system state space based on feature space transformation expression usingphase space reconstruction technology. Through time series characteristics of theforeign exchange market, we analysis of similarity of time series and use affine spread asemi-supervised clustering method to forecast the time series. At the end we establishnearest neighbor predictions.5．The model is tested using the recent10-year daily data of EUR/USD. In the end,we use mean square error and hit rate to measure the prediction accuracy. The test resultshows that the model has exhibited high prediction accuracy, thus supporting thevalidity of the model.