Simple Persistence Outperforms Machine Learning in Forecasting Thai Tourism Demand

In a study analyzing provincial monthly tourism demand across Thailand, researchers found that a straightforward historical persistence baseline outperformed complex machine-learning models during an untouched post-crisis test period. The empirical investigation evaluated visitor volume, tourism revenue, and tourist stay ratios across all 77 Thai provinces from January 2019 to February 2023. While sophisticated algorithms such as Random Forest, LightGBM, and Gradient Boosting were selected using rigorous time-aware cross-validation, their predictions yielded larger errors than simply assuming the previous month’s recorded values would persist into the next. This striking empirical divergence challenges the common assumption that higher algorithmic complexity inherently delivers superior predictive performance in volatile travel systems recovering from major public health disruptions.

Smart destination governance increasingly relies on forward-looking estimates of visitor flows, accommodation pressure, and regional revenue generation to optimize infrastructure, municipal staffing, pricing strategies, and marketing campaigns. Modern tourism systems do not behave as static continuations of past volumes; rather, they reflect dynamic interactions driven by seasonality, geographical accessibility, macroeconomic shifts, digital search interactions, and systemic health crises like the COVID-19 pandemic. Although tourism forecasting literature has steadily shifted toward complex multivariate machine-learning, deep neural networks, and hybrid architectures, these sophisticated frameworks often risk data leakage, temporal overfitting, and opaque decision logic. Model evaluations in the literature frequently present overly optimistic accuracy scores because feature extraction, hyperparameter tuning, or algorithm selection incorporate retrospective information unavailable at the true operational forecast date. The research team focused specifically on establishing an auditable, leakage-free forecasting architecture capable of handling moderate administrative panels characterized by extensive spatial breadth but relatively limited chronological depth.

The empirical investigation examined the official Thailand Domestic Tourism Statistics panel, which compiles monthly administrative records across all 77 provinces and five distinct administrative regions over 50 consecutive months, spanning January 2019 through February 2023. The raw administrative records comprised 30,800 long-format entries tracking eight primary tourism indicators per location and time step. To facilitate supervised machine-learning workflows, the team transformed these long-format records into a structured wide format of 3,850 province-month observations where each row captured a unique geographical unit in a given month. The analyzed timeline captured three distinct macroeconomic regimes: the pre-pandemic operating environment of 2019, the severe structural disruptions caused by pandemic containment measures from early 2020 through late 2021, and the initial tourism recovery phase extending from January 2022 through early 2023. This panel structure provided substantial spatial diversity across disparate urban and rural economies alongside rapid temporal regime transitions, establishing a demanding benchmark for predictive models.

To enforce methodological reproducibility and prevent predictive data leakage, the authors implemented strict information-set controls across four sequential operational tiers. All observations were sorted chronologically within each province before constructing any temporal or behavioral predictors, ensuring that no target observation could influence its own historical summary. The engineering pipeline constructed 183 distinct predictors from calendar trends, spatial identifiers, historical lag structures, rolling statistical windows, and regime indicators, while strictly excluding all contemporaneous, same-month tourism metrics from the input feature matrix. Constructing a 12-month historical memory requirement necessitated removing the initial 924 province-month target rows, leaving an active modeling panel of 2,926 observations. Lagged features incorporated values at one, two, three, six, and twelve months prior, capturing short-term persistence, half-year transitions, and annual seasonality. Rolling means and sample rolling standard deviations spanned past windows of three, six, and twelve months, shifted by one period so that every metric concluded strictly at the prior month.

The feature engineering framework incorporated cyclical calendar transformations and geographical indicators alongside historical behavioral metrics. Cyclical seasonality was captured using sine and cosine transformations of the calendar month number across an annual 12-month cycle, allowing algorithms to recognize numerical continuity between December and January. Two binary regime variables tracked structural macroeconomic conditions: an indicator for the active pandemic window from March 2020 through December 2021, and a subsequent indicator denoting the broader recovery phase starting in January 2022. Spatial heterogeneity was encoded through one-hot binary vectors representing all 77 unique provinces and five broader administrative regions, avoiding false ordinal assumptions across locations. The forecasting objective was formulated as three separate supervised regression problems targeting distinct operational dimensions of travel demand: total tourist arrivals, gross tourism revenue, and the tourist stay ratio. Rather than treating visitor volume as a proxy for all travel outcomes, each dependent variable was modeled independently to accommodate divergent physical, monetary, and duration dynamics across regional travel markets.

To explore a diverse spectrum of inductive biases, variance properties, and computational profiles, the researchers evaluated eight distinct regression methodologies alongside non-tuned historical rules. The evaluated algorithm library included an unweighted training mean baseline, standard ordinary least squares linear regression, L2-regularized Ridge regression, bagged decision tree ensembles using Random Forest and Extra Trees, sequential Gradient Boosting, and modern histogram-based boosting frameworks via XGBoost and LightGBM. Rather than employing randomized cross-validation that shuffles temporal sequences, the study enforced an expanding-window validation protocol restricted entirely to the pretest chronological block spanning January 2020 through June 2022. Across three sequential temporal folds, training windows progressively expanded from 18 to 26 months, evaluating models on out-of-sample blocks of four consecutive months using root mean squared error as the selection metric. The final eight months, extending from July 2022 through February 2023 and encompassing 616 province-month observations, remained completely untouched throughout hyperparameter tuning and model family selection.

However, the non-tuned previous-month persistence rule demonstrated superior point accuracy across all three targets during this untouched recovery window. In predicting the tourist stay ratio, persistence produced an R² of 0.8124 and an error of 5.7345 points, whereas Gradient Boosting recorded 8.8649 points. In contrast, an annual seasonal naïve baseline performed poorly throughout the evaluation, reflecting profound structural deviations between pre-crisis demand patterns and the subsequent travel recovery.

When researchers extended evaluation across multi-step forecasting horizons using direct estimation from rolling calendar origins, predictive accuracy degraded systematically as the lead time expanded from one to eight months ahead. For the pretest-selected models, mean root mean squared error for tourist arrivals climbed from approximately 80,612 at a one-month horizon to 386,391 at an eight-month horizon, while revenue error rose from 1.86 to 7.26 billion Thai Baht and stay ratio error expanded from 6.37 to 18.23 points. Residual diagnostics highlighted severe scale-dependent error concentration, with major travel hubs like Bangkok accounting for a disproportionate share of aggregate absolute forecast discrepancies. Formal non-parametric hypothesis testing using matched rolling-origin blocks revealed significant omnibus differences among the eight regression candidates via Friedman tests across all targets. However, subsequent pairwise Wilcoxon signed-rank tests adjusted with the Holm procedure yielded no statistically significant superiority for the pretest-selected models over their leading regression alternatives, demonstrating that fine-grained performance differences between competing algorithms could not be conclusively separated from temporal variation.

To prevent the deployment of opaque black-box systems in destination governance, the authors analyzed model reasoning using normalized global feature importance and local TreeSHAP explanations. Global attributions confirmed that short-term behavioral history dominated model predictions: one-month lagged total and domestic visitor volumes drove the arrival model, recent revenue and temporal indicators guided revenue predictions, and prior stay ratios dominated the stay ratio regressor. Local SHAP diagnostics for July 2022 illustrated divergent local mechanisms between destination tiers. In high-volume Bangkok, which recorded 3,322,469 actual arrivals against a prediction of 2,761,507, lagged domestic and total demand contributed massive positive adjustments, though the model still exhibited substantial underprediction. In contrast, for Sakon Nakhon, a typical-demand province near the national median with 128,451 actual arrivals and 121,479 predicted, positive and negative historical contributions largely balanced out. The authors explicitly emphasized that these game-theoretic attributions describe model mechanics rather than causal policy effects, cautioning destination authorities against interpreting feature weights as levers for municipal intervention.

Transforming predictive outputs into actionable destination management requires coupling numerical point forecasts with transparent reliability indicators, error history, and local behavioral attributions on interactive administrative dashboards. Computational assessments confirmed that all evaluated linear and tree ensemble models were lightweight and fast to train, with regularized and boosted models fitting within seconds, indicating that computational overhead is negligible compared to data governance and pipeline monitoring. However, the researchers emphasized that the framework’s empirical conclusions are bounded by important systemic constraints: the panel terminates in February 2023, capturing only early recovery dynamics following an unprecedented global shock. Furthermore, empirical findings derived from Thailand’s administrative geography cannot be assumed to generalize across differing international travel hierarchies, holiday schedules, or visa frameworks. The primary methodological contribution remains an auditable, leakage-free forecasting architecture that prioritizes rigorous temporal validation and humble baseline benchmarking over uncritical claims of algorithmic supremacy.

Subject of Research: Machine learning and benchmark validation for provincial tourism demand forecasting

Article Title: Smart tourism demand forecasting using machine learning and behavioral travel data

Article References: Wu, B., Wang, T., Yang, G., Xiong, X., & Chen, X. (2026). Smart tourism demand forecasting using machine learning and behavioral travel data. Discover Artificial Intelligence, 6(1), Article 1638. https://doi.org/10.1007/s44163-026-02621-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02621-1

Keywords: tourism demand forecasting, machine learning, smart tourism, time series validation, data leakage prevention, SHAP explanations, persistence benchmark, Thailand tourism, Smart, tourism, demand, forecasting