Sri Lanka Tourism Demand Forecast
Monthly tourist arrivals modelled from official SLTDA statistics, with the forecast model's error rate published so you can judge the numbers rather than take our word for them.
258,928
arrivals in Dec 2025, year on year
2.36M
total arrivals vs the previous 12 months
2.27M
SVR forecast, not a guarantee
8.7%
SVR, backtested. Lower is better
Monthly Tourist Arrivals
Official SLTDA actuals + SVR model forecast
Top Markets
By source country · Jan 2025 - Dec 2025
Share of total arrivals, with year-on-year change. Source: SLTDA arrivals by country.
How we built this
The method, the models we rejected, and what this forecast cannot tell you.
Data and features
We train on the official monthly arrivals series published by the Sri Lanka Tourism Development Authority. Nothing is scraped, and nothing is lifted from a press release. From that series we derive:
- Six months of arrival history (log scale), so the model learns momentum rather than raw magnitude
- Cyclical month encoding, so December sits next to January instead of eleven steps away
- Shock dummies for the 2019 Easter attacks, COVID-19, and the 2022 economic crisis
Those three shocks matter more than they look. A model that has not been told about them reads the 2020 collapse as seasonality and never recovers.
Models we tested
Five candidates, scored on a rolling-origin backtest over 222 forecasts — the model only ever predicts months it has not seen. We report MASE as the headline: below 1.0 means it beats a "same month last year" guess.
| Model | MAPE | MASE |
|---|---|---|
| SVRIn use | 8.7% | 0.35 |
| XGBoost | 14.1% | 0.57 |
| RandomForest | 15.7% | 0.66 |
| SARIMA | 65.5% | 3.02 |
| MLP | 499% | 20.12 |
SARIMA, the classical baseline, fails badly here — it cannot absorb three structural shocks in six years. The neural net (MLP) diverges outright on a series this short. We publish both rather than quietly dropping them.
What this cannot tell you
- A forecast is an estimate, not a promise. Ours is wrong by 8.7% on average, and it will be wrong by more than that when something genuinely new happens.
- No model anticipates a shock. Every one of the three we corrected for was invisible the month before it landed.
- SLTDA does not publish province-level monthly arrivals, so we show no regional split. We would rather show nothing than invent it.
Built and maintained by Ashan Lokuge. Found an error in the method? Tell us and we will publish the correction.