LAG 1 ONLINE CHANGEPOINT DETECTION WITH BAYESIAN INFERENCE

• Main Author: AZMIRA (NIM: 10114027), RIZKA
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/30566
• Institution: Institut Teknologi Bandung

Stationarity is a key assumption in sequential data analysis, the violation of which would render many prediction and estimation models inaccurate. In numerous cases, a process’ nonstationarity can be characterized by changepoints, which are points in the sequence where parameters of the underlying distribution change abruptly. The identification of these changepoints would mean that only data from the most recent changepoint was relevant to the analysis, and stationarity could be reassumed of this recent subset of data. This writing presents a model to detect changepoints in sequential data by observing a variable named run length. Evaluation of the occurrence of change is done at every point in the sequence by estimating the posterior distribution of run length conditioned on the sequence of observations. This conditional probability value is obtained recursively by Bayesian inferential methods, such that the model is computationally efficient and is suitable for analysis on an ongoing data stream. The model is further enhanced by the addition of time lag, which is shown to improve the model both in terms of precision and sensitivity. Application is made to two datasets: 1) the daily returns of JKSE and 2) black and white images for edge detection purposes.