An Integrated Software Package for Econometric and Time-Series Analysis and Modeling: Multimodeler

Full abstract

Read the full abstract

This paper describes Multimodeler software which is a comprehensive software suite for econometric analysis, time-series modeling, forecasting, and comparative model evaluation. Multimodeler is designed as an integrated graphical environment, it implements a broad collection of classical econometric methods, statistical time-series techniques, machine-learning approaches and neural-network models within a single workflow. Multimodeler is intended to enable researchers, students, and practitioners to move from data import and exploratory analysis to model estimation, diagnostic testing, forecasting and performance comparison among classical and neural network models without relying on multiple independent software packages or extensive programming. Multimodeler incorporates descriptive statistics, correlation and covariance analysis, graphical visualization, frequency-domain analysis, filtering, detrending, differencing, seasonal adjustment, and outlier detection. Stationarity and unit-root analysis is supported through Augmented Dickey-Fuller (ADF), Phillips-Perron (PP), and KPSS procedures with configurable deterministic components, lag selection, and bandwidth options. The modeling environment includes linear regression; AR, MA, ARMA, ARIMA, SARIMA, ARIMAX, and SARIMAX models; exponential-smoothing methods; ARCH and GARCH volatility models; VAR, VARMA, and VARMAX frameworks; Engle-Granger cointegration analysis; error-correction models; Johansen-based VECM analysis; and state-space and Kalman-filtering methods. Both ordinary least squares and maximum-likelihood estimation approaches are available for relevant model classes. Moreover, nonlinear models such as Cobb-Douglas modeling, logistic function modeling and Michaelis-Menten type modeling are also included. Multimodeler also incorporates multilayer perceptron based artificial neural-network and deep-learning-based modeling and forecasting architectures including autoregressive, exogenous-input, and autoregressive-with-exogenous-input configurations. Neural models can be trained using conventional backpropagation-based optimization, momentum methods, Adam optimization, and population-based Artificial Bee Colony optimization. Manual and automated model-search facilities allow alternative lag structures, network architectures, activation functions, and parameter configurations to be systematically evaluated within a single button click. In Multimodeler, model performance assessment is supported by a wide range of goodness-of-fit, diagnostic, and forecasting measures, including RMSE, MAE, MAPE, sMAPE, MdAPE, MASE, normalized error measures, R², adjusted R², information criteria, residual diagnostics, autocorrelation analysis, and Ljung-Box testing. Automated model-selection procedures and comparative evaluation tools are intended to make it possible to examine competing econometric and computational models within a consistent analytical framework rather than treating each methodology as an isolated procedure. This paper presents the design philosophy, methodological foundations, software architecture, principal modules, estimation procedures, diagnostic tools and forecasting capabilities of Multimodeler. Particular emphasis is placed on its role as a unified environment in which classical econometric models, modern time-series methods, and neural forecasting approaches can be estimated, validated and compared using common data and common performance criteria. Multimodeler is positioned as both an applied research platform and an educational tool for systematic, transparent and reproducible econometric and time-series analysis, modeling and forecasting.