📊 Tourist Arrivals Forecast

Data source: Real model output — outputs/dashboard_data.json · Last run: —

Production Ready

ARIMAX

Primary forecasting · Revenue & capacity planning · Daily ops

Uncertainty Expert

BSTS

Risk management · Scenario analysis · Confidence intervals

Monitoring Tool

Prophet

Anomaly detection · Holiday effects · Trend break alerts

Cross-validation MAPE — lower is better

ARIMAX

—

Prophet

—

BSTS

—

Industry benchmark: 10–15% · Evaluation: rolling CV h=1..6 months

Key metrics comparison

Metric	ARIMAX	BSTS	Prophet	Benchmark
CV MAPE	—	—	—	10–15%
Retrain speed	2s	45min	15s	<30s
CI coverage	—	—	—	>75%
Status	Primary	Secondary	Monitor	—

Best model MAPE

—

Data range

—

Peak month (ARIMAX)

—

Annual total (ARIMAX)

—

Projected 2026

Toggle:

MAPE by forecast horizon

ARIMAX BSTS Prophet Benchmark 15%

RMSE by forecast horizon (log scale)

ARIMAX BSTS Prophet

Compute performance benchmark

Model	Training time	Prediction time	Key dependencies	Retraining effort	Interpretability
ARIMAX	~2s	0.1s	forecast, tseries	Low	High
Prophet	~15s	1s	prophet (Stan)	Medium	High
BSTS	~45min	30s	bsts (MCMC)	High	Medium

Overall MAPE ranking (real CV results)

Revenue calculator:

ADR (USD/night)

Avg stay (nights)

—

ARIMAX revenue

—

BSTS revenue

—

Prophet revenue

Scenario:

Monthly arrivals forecast — Jan to Dec 2026 (real model output)

Month	ARIMAX	ARIMAX Lo 95%	ARIMAX Hi 95%	BSTS median	BSTS Lo 95%	BSTS Hi 95%	Prophet

Peak month (ARIMAX)

—

Airport monthly cap

350k

Maximum capacity

Months exceeding cap

None

All months within capacity

All-model forecast + historical (real data)

HistoricalARIMAXBSTSProphet

Month-over-month growth % (ARIMAX)

Primary ARIMAX — Daily operations —

Revenue forecastingCapacity planningHotel bookingsAirport schedulingRegulatory reportingBudget allocation

Best accuracy. Retrains in 2 seconds. MAPE: —

Secondary BSTS — Risk management —

Uncertainty bandsStress testingInsurance pricingScenario planningStakeholder CIsFinancial planning

Use when uncertainty quantification needed. CI coverage: —. MAPE: —

Monitoring Prophet — Anomaly detection —

Shock detectionHoliday effectsTrend breaksBackup validationSeasonal patterns

Optional — may skip if resources are limited. MAPE: —

Confidence scores (derived from MAPE, CI coverage, retraining speed)

ARIMAX—

BSTS—

Prophet—

Score = 0.5×(norm MAPE) + 0.25×(CI coverage) + 0.25×(norm speed) · Based on real CV metrics

Decision matrix

Use case	ARIMAX	BSTS	Prophet
Daily revenue forecast	Best	Slow	OK
Uncertainty intervals	None	Best	Limited
Shock/crisis events	Good	Good	Good
Holiday effects	Manual	Manual	Auto
Retraining speed	2s	45min	15s
Interpretability	High	Medium	High
CI coverage	—	76.9%	—

MAPE heatmap — model × horizon (real CV results)

Auto-retraining status

Last run: — · Next SLTDA data upload → re-run main.R → JSON auto-updates

ARIMAX residuals — white noise check (simulated pattern)

CV prediction error spread by model (real CV data)

ARIMAX diagnostics

Ljung-Box p > 0.05 — white noise residuals

Model: —

Shock dummies significant

Manual regressor updates needed on retrain

Prophet diagnostics

Logistic growth converged (cap 350k)

Holiday effects: Vesak, Sinhala New Year

aes_string deprecation warning (suppressed)

Timezone fix applied

BSTS diagnostics

Local linear trend + seasonal converged

1,000 MCMC iterations (burn = 200)

CI coverage: —% (target >75%)

Slow retraining (~45min)

Data coverage

— · — obs

Easter 2019 shock dummy coded

COVID 2020–2021 coded

Economic crisis 2022 coded

Cross-validation summary (real results)

Model	CV method	Avg MAPE	Avg sMAPE	RMSE (log)	MAE (arr)	95% CI coverage	Recommend retrain