---
title: "A Review of ARIMA vs. Machine Learning Approaches for Time Series Forecasting"
authors: "(Synthesis: MDPI Future Internet review + Siami-Namini & Namin ICMLA 2018)"
source: "MDPI Future Internet / IEEE ICMLA"
source_url: "https://www.mdpi.com/1999-5903/15/8/255"
paper_id: "MDPI Future Internet 15(8):255 (2023); Siami-Namini & Namin (2018)"
harvested: 2026-08-22
year: 2023
keywords: [ARIMA, LSTM, machine learning, time series, forecast comparison]
---

## Abstract / Summary

The ARIMA-vs-machine-learning literature (MDPI review 15(8):255, 2023; Siami-Namini & Namin,
ICMLA 2018) compares classical ARIMA forecasting with machine learning / deep learning models
(notably LSTM) across many time-series tasks. The consolidated findings across dozens of
comparisons are consistent:
- **ARIMA** is simpler, theoretically grounded, interpretable, and efficient at **short-horizon**
  forecasting with modest data; its structure (lags via AIC/BIC) is well understood.
- **LSTM / ML models** tend to show a slight edge with **large datasets and longer horizons**, but
  are harder to interpret, have many more tuning parameters, and are prone to overfitting on small
  samples.
- On many tasks the two are **comparable**; ML's advantage is not universal and is data-dependent.

**Relevance to StratMac RE-sector work — a caution, not a path.** Most housing ML applications in
this space are shallow: a single fitted result, no walk-forward discipline, no no-lookahead lag
alignment, small N, and often no reported native-unit error. Per StratMac's marginal-value
discipline, do **not** be pulled into a headline ML/LSTM/boosting model. If ML is pursued at all, it
belongs strictly as a **benchmark-to-beat** (e.g. gradient boosting on the lag-aligned matrix) that
must survive walk-forward AND beat the parsimonious ridge before it earns a result — the same
bar any candidate feature must clear. LSTM's large-data edge is precisely the regime StratMac's
long-horizon design avoids (thin recession/housing-turning samples).