Model-free prediction and regression. A transformation-based approach to inference (Q269798)

The author presents model-free prediction and regression as opposed to model-based ones. The prediction task is presented by the following regression setup. A vector of observed responses \(Y_n=(Y_1,Y_2,\ldots,Y_n)'\) is given which is associated with the vector of regressors \(X_n=(X_1,X_2,\ldots,X_n)'\) and an unobserved value \(Y_{n+1}\) corresponding to \(X_{n+1}\) (a regressor value of interest) should be predicted. If \(Y_i\) are i.i.d., the MSE-optimal predictor of \(Y_{n+1}\) is simply given by the expected value of \(Y_i\; (i=1,\ldots,n)\). In the case where the i.i.d. assumption breaks down, model-free prediction principle uses the structure of the problem in order to find an invertible transformation \(H_m\) that can map the non-i.i.d. vector \(Y_m\) to a vector \(\epsilon_m(\epsilon_1,\ldots,\epsilon_m)'\) that has i.i.d. components. Here \(m\) could be taken equal to either \(n\) or \(n+1\) as needed. In this way, the prediction problem is reduced to the trivial one of predicting i.i.d. variables. The book is organized in four parts. The first part (Part 1, The model-free prediction principle) presents the model-free prediction principle. The second part (Part 2, Independent data: regression) considers model-based prediction in independent-data regression (Chapter 3, Model-based prediction in regression), model-free prediction in independent-data regression (Chapter 4, Model-free prediction in regression), and compares model-free with model-based confidence intervals for independent data (Chapter 5, Model-free vs. model-based confidence intervals). In Part 3 (Dependent data: time series) optimal linear prediction for linear time series is considered (Chapter 6, Linear time series and optimal linear prediction), detailed construction of prediction intervals in model-based regression is described (Chapter 7, Model-based prediction in autoregression). Chapter 8 (Model-free inference for Markov processes) investigates model-free autoregression, i. e., Markov processes, to serve as a contrast to the model-based autoregression of Chapter 7. In Chapter 9 (Predictive inference for locally stationary time series), a locally stationary model is considered where stochastic structure of time series is slowly-changing. Prediction is performed based on decomposition of the time series into three components: trend, seasonal and stationary part. In Part 4 (Case study: model-free volatility prediction for financial time series), model-free and model-based volatility prediction are compared on three representative datasets of daily returns taken from a foreign exchange rate, a stock price and a stock index. The monograph restores the emphasis on observable quantities. Considering model-free and model-based prediction, the monograph emphasises on model-free approach but also shows the close relation between these two approaches. The book is of interest for both academics and practitioners in the field of data analysis.