Short Time Series Forecasting: Recommended Methods and Techniques

Abstract

This paper tackles the problem of forecasting real-life crime. However, the recollected data only produced thirty-five short-sized crime time series for three urban areas. We present a comparative analysis of four simple and four machine-learning-based ensemble forecasting methods. Additionally, we propose five forecasting techniques that manage the seasonal component of the time series. Furthermore, we used the symmetric mean average percentage error and a Friedman test to compare the performance of the forecasting methods and proposed techniques. The results showed that simple moving average with seasonal removal techniques produce the best performance for these series. It is important to highlight that a high percentage of the time series has no auto-correlation and a high level of symmetry, which is deemed as white noise and, therefore, difficult to forecast.

Forecasting the Propagation from Meteorological to Hydrological and Agricultural Drought in the Huaihe River Basin with Machine Learning Methods

“…Kim et al [89] discussed the better forecasting performances for streamflow forecasting in the high flow (higher autocorrelation coefficients) regime than those in the low flow regime. The data with no autocorrelation, deemed as white noise, is difficult to forecast [90].…”

Hemoglobin signal network mapping reveals novel indicators for precision medicine

“…It is likewise appreciated that well-established non-network time-series methods, while useful for classification 21 or feature-extraction 22 – 24 , and potentially useful for pre-conditioning operations for the method of this report, do not generate data from which low-variance measures of short-term dynamics can be directly read. Additionally, while these methods lie on a spectrum with regard to the requisite degree of prior knowledge of the processes governing system dynamics (e.g., low for wavelets 21 , 22 , high for Kalman filtering and related methods 25 , 26 ), mathematically they are more strongly assumption-dependent or model-based than the fr-OPN method used here.…”

Impact of Xpert MTB/RIF implementation in tuberculosis case detection and control in Brazil: a nationwide intervention time-series analysis (2011–2022)

“…Next, to perform an ITSA, we use the intervention year detected as a level shift in the model to segment our data into pre- and post-intervention phases. Following previously published studies for accurately forecasting small-sized time-series, 12 , 13 we then employed ARIMA and ForecastHybrid models 12 , 13 to predict the scenario without Xpert adoption (counterfactual scenario) from data collected prior to its implementation and compared their accuracy measures with Mean Absolute Percentage Error (MAPE). Afterwards, the most accurate predictions of the hypothetical (counterfactual) scenarios were then compared to the real (observed) values post-intervention to assess the impact of Xpert implementation.…”