Evaluating the performance of long short-term memory to other time series models for univariate and multivariate data

Time series data analysis is one of the difficult tasks in real-world data problems because the data entities are determined not only by physical causes, but also by the chronological sequence in which they occurred. In this work, we have examined how Long Short-Term Memory (LSTM) model is perform well to traditional statistical methods and other machine learning models for performing univariate and multivariate time series single-step forecasting. We have also studied how memory cells and gated structures work in LSTM model with different cases to show advancement of LSTM as compared to other time series analysis tools. The key contributions of this work are methodological and empirical rather than architectural. Specifically, this study (i) provides a mechanistic, architecture-level interpretation of LSTM’s gating and memory dynamics that explains its temporal learning behavior, distinct from post-hoc explainable AI (XAI) feature-attribution techniques, and (ii) establishes a benchmarking framework that evaluates LSTM against statistical models, machine learning models, a forecasting-specific Transformer architecture (PatchTST), and a baseline (vanilla) Transformer under identical experimental settings on a real-world environmental dataset. Experimental results demonstrate that, within context-specific empirical evidence, the LSTM model exhibits more stable and consistent performance than the considered baselines. Cross-regional validation further shows that the observed performance trends remain largely consistent across both datasets, suggesting that the advantages of LSTM are not solely attributable to local characteristics of a single region. The findings are presented as context-specific empirical evidence rather than claims of universal model superiority and provide practical insights into forecasting under data-constrained environmental conditions.