## Modeling and Forecasting Monthly Tourism Arrivals to Aruba Since COVID-19 Pandemic — Executive Summary

## Source details

**Canonical URL:** [Modeling and Forecasting Monthly Tourism Arrivals to Aruba Since COVID-19 Pandemic — Executive Summary](https://www.imf.org/-/media/files/publications/wp/2022/english/wpiea2022226-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2022/english/wpiea2022226-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2022/english/wpiea2022226-print-pdf.pdf.json)

---

### Pandemic impact and tourism recovery
- Aruba closed its borders for inbound travelers for almost three months in early 2020 and began welcoming tourists back in June 2020.
- Aruba received about 34 percent and 72 percent of the 2019 tourist inflows in 2020 and 2021, respectively.
- In 2020, tourism arrivals to Aruba dropped by about 2/3 of the 2019 level, triggering a shrinkage of real economy by 18.6 percent.
- In 2021, stay-over arrivals more than doubled, reaching 72 percent of 2019 level. In 2022, recovery continued and in April-June arrivals already exceeded the 2019 levels in the same months.
- COVID-19 pandemic reduced total tourism arrivals to Aruba by an average of 43-44 thousand tourists in 2020, and by 38-39 thousand visitors over 2020-2021.

### Forecasting goal, modeling approach, and samples
- Objective: Improve modeling and forecasting of monthly tourism arrivals to Aruba since the pandemic using variables capturing structural and behavioral changes and available for forecasting.
- Primary framework: ARIMA (simplified to ARMA since d=0) and ARIMAX with monthly dummies; seasonal binary indicators used (11 dummies, dropping September).
- Main ARIMA specifications considered: (p=1, q=1), (p=2, q=1), (p=2, q=2); preferred full model: ARIMA (2,0,2) with exogenous regressors.
- Full model (example equation [6]/[7]) includes two AR terms, two MA terms, 11 monthly dummies, and four exogenous regressors: d1RPCE t−3, GT t−2, CAP t, d1VIX t; plus optional intercept correction IC (ICovid dummy value = 1 from 2020m4-2023m12).
- Estimation samples:
  - Original pre-pandemic estimation sample: 2004m1–2019m12 (192 monthly observations).
  - Extended samples: 2004m1–2020m12 and 2004m1–2021m12.
  - Full sample used in paper: 2004m1–2022m6 (222 monthly observations).
- Forecast horizon: out-of-sample 18 months ahead (2022m7–2023m12); short-term horizons h=0-3 months emphasized.

### Data, timing, and stationarity
- Dependent variable: monthly tourist arrivals to Aruba (in thousands of persons) from CBA monthly tables, cross-checked with ATA.
- Main explanatory series:
  - US Real personal consumption expenditures per capita (RPCEcap) — third lag of first difference used: L3.dRPCE.
  - US Real personal income per capita (RPIcap) — first differences used when nonstationary: L3.dRPI.
  - Google Trends for “Aruba all-inclusive hotels” (GT), lagged two months (GTt−2).
  - Flight capacity utilization ratio (CAP) from US BTS (percent, 0–100), concurrent.
  - Chicago Board Options Exchange Volatility Index (VIX) — monthly maximums, used as first difference (d1VIX).
  - Robustness: US EPU1, EPU2, and US unemployment rate (UR).
- Timing of data releases: US income/expenditure/population series lag about 2 months; Aruba tourism arrivals lag about 3 months; Google Trends and VIX available almost in real time (few days delay, up to one month for monthly aggregates); US airlift data delayed 6-7 months.
- Stationarity (selected ADF/DFGLS results, N=192 unless noted):
  - ARR: Mean=77.12, St.d.=15.63, Min=48.00, Max=118.47; ADF=-4.49***, DFGLS=-2.48* (no unit root).
  - CAP: Mean=80.99, St.d.=6.25, Min=58.89, Max=92.46; ADF=-5.39***, DFGLS=-1.73** (no unit root).
  - GT: Mean=26.91, St.d.=10.15, Min=2.44, Max=62.88; ADF=-8.30***, DFGLS=-6.69*** (no unit root).
  - RPIcap: Mean=44.83, St.d.=3.02, Min=40.01, Max=51.02; ADF=-0.24, DFGLS=2.03 (unit root not rejected → use first difference). D1.RPIcap: Mean=0.057, St.d.=0.32; ADF=-15.88***, DFGLS=-11.39*** (stationary).
  - RPCEcap: ADF=1.10, DFGLS=4.70 (unit root not rejected → use first difference).
  - VIX: Mean=23.50, St.d.=11.61; ADF=-4.92***, DFGLS=-4.09*** (stationary in levels; differenced in regressions).
  - UR: Mean=6.07, St.d.=1.93; ADF=0.09, DFGLS=-0.31 (unit root not rejected → use first difference).
- Unit-root tests used: ADF and DFGLS; variables with unit roots entered as first differences.

### Key predictors, structural changes, and behavioral assumptions
- Market focus: US visitors account for over 80 percent of all tourists → models concentrate on US-related explanatory variables.
- Shift in predictor importance post-pandemic:
  - RPCE per capita gained predictive importance over RPI per capita because pandemic fiscal transfers boosted RPI but did not translate proportionally into spending.
- Robust and consistently significant predictors across samples and specifications:
  - Google Trends (GT) for “Aruba all-inclusive hotels” — always significant and strong predictor.
  - Flight capacity utilization (CAP) — always significant and strong predictor.
- VIX and uncertainty measures:
  - Inclusion of the VIX improves models before intercept correction; VIX loses significance once a COVID-19 binary indicator is included.
  - Alternative risk measures (EPU1, EPU2, UR) do not materially improve model fit relative to VIX.
- Lag structure and consumer planning:
  - Consumers start planning about three months in advance: US income/expenditure enter with third lag; GT lagged by two months; CAP concurrent; VIX concurrent (first difference); EPU/UR lagged one month when used.

### Main empirical results and model performance
- Preferred baseline: ARIMA (2,0,2) with exogenous RPCE, CAP, GT, and VIX yields best fit and lower forecast error for full sample; when sample includes 2021 ARIMA (1,0,1) can produce lower AIC in some cases.
- Coefficients and marginal effects (selected, preserve sample context):
  - CAP coefficient progression: 0.37*** (2004m1-2019m12); 0.60–0.61*** (2004m1-2020m12); 0.71*** (2004m1-2021m12).
    - Based on the full sample, one more percent in load factor would increase the tourism inflow to Aruba by about 710 persons.
  - GT (L2.GT) coefficients: 0.09** / 0.08* / 0.10** / 0.13** across specifications and samples.
    - An increase of internet search frequency by one percent corresponds to 130 additional visitors (based on the full sample).
  - RPCE (L3.dRPCE) coefficients (example ranges): 1.72, 1.66, 0.85 (pre-pandemic) and 1.33*, 1.27*, 1.25* (with 2020) and 1.35*, 1.26*, 1.20* (with 2021).
    - Additional 100 USD per capita real spending in the USA would generate about 125 more tourist arrivals to Aruba (when pandemic months are included).
  - VIX (D.vixmax) coefficients vary and weaken after IC: e.g., -0.05, -0.09***, -0.02 across samples/specs; pre-pandemic a 1-point increase in max monthly VIX cost Aruba about 5-9 tourists per month; extended samples and IC reduce coefficient to -2 and lose statistical significance.
- Model fit and diagnostics (selected figures):
  - SD (selected): 3.98***–4.11*** (2004m1-2019m12); 4.20***–4.25*** (through 2020); 4.50***–4.54*** (through 2021).
  - LL examples: -533.7, -533.2, -528.8 (pre-pandemic); -571.8, -570.5, -569.4 (through 2020); -620.2, -619.3, -618.4 (through 2021).
  - AIC examples: 1097.7–1106.4 (pre-pandemic); 1182.8–1183.6 (through 2020); 1280.4–1280.8 (through 2021).
  - SWILK: 0.99** / 0.99 across specifications; residual normality rejected at 10 percent in some pre-pandemic samples.
- Variables’ contributions to model improvement (reported shares):
  - RPCE: contributes more than 70 percent to LL improvement; 14.2 percent to AIC improvement; 2.1 percent to residual reduction.
  - CAP: largest role in decreasing the error term and AIC.
  - Intercept Correction (IC): contributes 16.5 percent to error reduction, 2.2 percent to higher LL, and 17 percent to lower AIC.
  - VIX: contributes 5.1 percent to error reduction, 4.3 percent to LL improvement, and 3.9 percent to AIC improvement.
  - Without IC, four exogenous variables reduce residuals by 29 percent, improve LL by 10 percent, and reduce AIC by 11 percent (vs pure ARIMA). After IC, improvements are 34, 13, and 14 percent respectively.

### Forecast accuracy, RMSE effects, and out-of-sample forecasts
- RMSE (estimation sample: 2004m1–2021m12) — ARIMA 202 examples:
  - Full sample RMSE by nested models: None = 7.24; RPCE = 7.14; RPCE, CAP = 5.17; RPCE, CAP, GT = 5.08; RPCE, CAP, GT, VIX = 4.99; RPCE, CAP, GT, VIX, IC = 4.69.
  - Training sample RMSE: None = 6.98; RPCE = 6.85; RPCE, CAP = 5.15; RPCE, CAP, GT = 5.06; RPCE, CAP, GT, VIX = 4.98; RPCE, CAP, GT, VIX, IC = 4.69.
  - Test sample RMSE: None = 13.77; RPCE = 13.85; RPCE, CAP = 5.66; RPCE, CAP, GT = 5.74; RPCE, CAP, GT, VIX = 5.23; RPCE, CAP, GT, VIX, IC = 7.25.
  - MA2 term contributes 95–98 percent of RMSE reduction across samples; CAP single most important factor in reducing RMSE.
- Out-of-sample 18-month forecast (model chosen: ARIMA 202 with RPCE, CAP, GT, VIX, no IC for reported baseline):
  - Average point forecast: 92.8 thousand tourist arrivals per month (2022m7-2023m12).
  - Lower bound (5 percent confidence level): 78.9 thousand tourists per month.
  - Upper bound (5 percent confidence level): 106.6 thousand tourists per month.
  - Cumulative/baseline scenario: 528.0 thousand tourists in 2022H2; 589.4 thousand in 2023H1; 623.4 thousand in 2023H2.
  - Overall indicated tourism growth of 5.3 percent in 2023.
- Forecast uncertainty: bounds obtained as point forecast ± RMSE × sqrt(forecast horizon).

### Robustness checks and sensitivity
- Alternative risk proxies (US unemployment rate UR, US EPU1, EPU2) and replacing RPCE with RPI:
  - Models replacing VIX with UR, EPU1, or EPU2 deteriorate model fit and forecast properties.
  - Models with RPI per capita instead of RPCE lose significance once 2020-21 are included and perform worse in extended samples.
- Annex 5 (selected robustness table highlights):
  - CAP remains consistently positive and highly significant across specifications (examples: 0.37***; 0.59***; 0.70***; 0.71***).
  - GT lagged retains positive significance across samples (L2.GT around 0.08–0.13 with * or **).
  - Intercept Corrections (IC) are large negative and highly significant (e.g., -43.90***; -38.97***; -39.94***).
  - EPU dummies reported near-zero coefficients; VIX small and mostly not significant.

### Policy and operational implications
- Accurate short-term tourism forecasts that incorporate GT and CAP and account for structural breaks are critical inputs for Aruba’s macroeconomic forecasting model ("MARUBA") where tourism inflows are among the most critical inputs for projecting RGDP.
- Recommended forecasting practice for analysts in small tourism-dependent economies:
  - Prioritize RPCE per capita over RPI per capita for near-term demand signals during and after pandemic-like shocks.
  - Incorporate real-time indicators such as Google Trends and flight capacity utilization to capture consumer intent and supply-side constraints.
  - Use intercept correction and pandemic indicators to account for structural shifts and normalize residuals, while testing out-of-sample impacts due to IC’s binary nature.
  - Recognize limitations of traditional variables (e.g., REER) for short-term forecast improvements.
  - Re-estimate models as new observations become available; including more observations generally improves forecast accuracy.

*wpiea2022226-print-pdf - Executive Summary*

### Executive Summary ......................................................................................................

### Executive Summary

### Document structure and sections
- Executive Summary (page 4)
- I. Introduction (page 5)
- II. Literature Review (page 7)
- III. Data: Sample, Sources, and Their Properties (page 8)
- IV. Methodology, Models, and Assumptions (page 10)
- V. Main Empirical Results: Estimated Models and Forecasts (page 14)
- VI. Conclusion (page 19)
- Annex 1. Notes on Google Trends Data (page 21)
- Annex 2. Panel Charts: Dynamics in Data (page 23)
- Annex 3. Main Emirical Results (page 25)
- Annex 4. Panel Charts: Projected Arrivals (page 30)
- Annex 5. Robustness Check (page 31)
- References (page 33)

### Figures listed in the content unit
- Figure 1. Tourist Arrivals to Aruba: Dynamics, Sources of Origin, and Macroeconomic Impact (page 5)
- Figure 2. Role of Regressors and Model Terms in the RMSE improvement (page 18)
- Figure 3. In-Sample Tourist Arrivals Projections (page 18)
- Figure A2-1. Dynamics in Tourist Arrivals and Explanatory Variables (page 23)
- Figure A4-1. Projected Arrivals: Comparison across ARIMA classes, before and after IC (page 30)

### Tables listed in the content unit
- Table 1. Summary Statistics and Data Properties (2004m1 – 2019m12) (page 10)
- Table 2. Tourist Arrivals to Aruba—Model w/RPCE, GT, CAP, VIX (with intercept correction) (page 15)
- Table 3. Tourist Arrivals to Aruba—ARIMA 202—Models’ Comparison (page 16)
- Table 4. Variables’ Role in Reduction of the Forecast Error (est. sample: 2004m1 – 2021m12) (page 17)
- Table A3-1. Tourist Arrivals to Aruba—Model w/RPCE, GT, CAP, VIX, IC and Monthly Dummies (page 25)
- Table A3-2. Tourist Arrivals to Aruba — Auxiliary Models w/RPCE (Sample: 2004m1-2021m12) (page 26)
- Table A3-3. Contributions of Exogenous Regressors to the Model Fit (page 28)
- Table A3-4. Forecast Evaluation — Contributions of Regressors to Forecast Accuracy (page 29)
- Table A5-1. Tourist Arrivals to Aruba—Model w/RPI, GT, CAP, VIX, IC and Monthly Dummies (page 31)
- Table A5-2. Tourist Arrivals to Aruba—Models with alternative risk measures (page 32)

### Core methodological and data elements (as listed)
- Model classes and methods referenced:
  - ARIMA (Auto Regressive Integrated Moving Average)
  - ARIMAX (ARIMA model with exogenous variables)
  - ARMA (Auto Regressive Moving Average)
- Statistical criteria and tests referenced:
  - ADF test (Augmented Dickey–Fuller test)
  - DFGLS (The modified Dickey–Fuller t test for a unit root in a time series)
  - AIC (Akaike Information Criteria)
  - RMSE (Root Mean Squared Error)
  - MAE (Mean Absolute Error)
  - Theil U (Theil Inequality Coefficient)
- Data sources and series abbreviations referenced:
  - RPCE Real Personal Consumption Expenditures
  - RPI (mentioned in Table A5-1 title)
  - GT Google Trends
  - CAP (capitalization or capacity variable as used in table titles)
  - VIX (volatility index)
  - REER Real Effective Exchange Rate
  - RGDP Real Gross Domestic Product
  - ATA Aruba Tourism Authority
  - BEA Bureau of Economic Analysis
  - BTS Bureau of Transportation Statistics
  - CBA Central Bank of Aruba
  - WHO World Health Organization
  - WTO World Tourism Organization
- Sample periods explicitly noted:
  - 2004m1 – 2019m12 (Table 1)
  - 2004m1 – 2021m12 (Tables and estimated sample references)

### Analytical focus and scope (as signposted)
- Modeling and forecasting monthly tourism arrivals to Aruba since the COVID-19 pandemic.
- Emphasis on:
  - Incorporating exogenous regressors (RPCE, GT, CAP, VIX, etc.)
  - Intercept Correction (IC) and monthly dummies in model specifications
  - Model comparison across ARIMA classes and evaluation of forecast accuracy
  - Assessing contributions of regressors to model fit and reduction in forecast error
  - Robustness checks and alternative risk measures

*Source: wpiea2022226-print-pdf - Executive Summary (IMF Working Paper).*

### Executive Summary

### Executive Summary

### Pandemic impact and tourism recovery
- The Covid-19 pandemic triggered containment measures and disrupted businesses worldwide, severely affecting tourism destinations. Aruba closed its borders for inbound travelers for almost three months in early 2020 and began welcoming tourists back in June 2020.
- Aruba received about 34 percent and 72 percent of the 2019 tourist inflows in 2020 and 2021, respectively.
- In 2020, tourism arrivals to Aruba dropped by about 2/3 of the 2019 level, triggering a shrinkage of real economy by 18.6 percent.
- In 2021, stay-over arrivals more than doubled, reaching 72 percent of 2019 level. In 2022, recovery continued and in April-June arrivals already exceeded the 2019 levels in the same months.

### Forecasting goal and modeling approach
- Objective: Improve modeling and forecasting of monthly tourism arrivals to Aruba since the pandemic, focusing on variables that capture structural and behavioral changes (including Google Trends and flight capacity data) and that are available historically and for forecasts.
- Primary modeling framework: ARIMA (simplified to ARMA since d=0) with monthly dummies; models tested over three samples and forecasts evaluated over 2020m1-2022m6.
- Original estimation sample: 2004m1-2019m12 (192 monthly observations). Full sample used in the paper: January 2004 to June 2022 (222 monthly observations).
- Forecast horizon: out-of-sample forecasts of tourism arrivals to Aruba 18 months ahead; short-term horizons considered useful (h=0-3 months) given flight capacity dynamics.

### Key predictors and structural changes since COVID-19
- Market focus: Given that visitors from the USA constitute over 80 percent of all tourists, the model concentrates on US-related explanatory variables.
- Relative predictor importance shifted post-pandemic:
  - Real Personal Consumption Expenditure (RPCE) became a more significant predictor than Real Personal Income (RPI). RPI was affected by pandemic fiscal support that did not translate into spending, whereas RPCE reflected mobility-constrained spending behavior.
- Robust and consistently significant predictors:
  - Google Trends for All-Inclusive Hotels in Aruba (GT) is always significant and a strong predictor across samples and specifications.
  - Flight capacity utilization (CAP, load factor) is always significant and a strong predictor across samples and specifications.
- Risk and uncertainty measures:
  - Inclusion of the VIX improves the model before intercept correction, but VIX loses significance once a binary indicator for the COVID-19 pandemic is included.
  - Model estimates are not sensitive to alternative risk measures tested: two variants of the US Economic Policy Uncertainty Index (EPU1, EPU2) and the US unemployment rate (UR).
- Other variables:
  - REER is dropped from the baseline specification because, while weakly significant in the estimation sample, it does not improve out-of-sample forecasting properties.
- Intercept correction:
  - Intercept correction helps to normalize residuals in the full sample and aids model performance post-structural shift.

### Data, stationarity, and timing
- Dependent variable: monthly tourist arrivals to Aruba (in thousands of persons) from CBA monthly tables, cross-checked with Aruba Tourism Authority (ATA).
- Sample span: 2004m1–2022m6 (222 observations); original pre-pandemic estimation sample 2004m1–2019m12 (192 observations).
- Important data features:
  - Aruba has well-defined seasonality with lowest arrivals in September; historical arrivals affected by external shocks (GFC, Venezuela crisis, COVID-19).
  - Timing of data releases: US income/expenditure/population series published with a lag of about 2 months; Aruba tourism arrivals published with a lag of about 3 months; Google Trends and VIX available almost in real time (few days delay, up to one month for monthly aggregates); US airlift data available with a delay of 6-7 months.
- Main explanatory series used:
  - US Real personal income (RPI) and Real personal consumption expenditures (RPCE) per capita (trillions of chained 2012 USD series from St. Louis Fed, transformed to per capita measures).
  - US carriers’ flight capacity utilization ratio (CAP) from US BTS: total passengers divided by available seats (percent, 0–100).
  - Google Trends (GT) for “Aruba all-inclusive hotels” from the USA.
  - Chicago Board Options Exchange Volatility Index (VIX) based on monthly maximums as proxy for financial risk/uncertainty.
  - Robustness checks: US EPU1, EPU2, and US unemployment rate (UR).
- Stationarity findings (selected):
  - Tourist Arrivals to Aruba (ARR): N=192, Mean=77.12, St.d.=15.63, Min=48.00, Max=118.47; ADF=-4.49***, DFGLS=-2.48* (no unit root).
  - CAP: N=192, Mean=80.99, St.d.=6.25, Min=58.89, Max=92.46; ADF=-5.39***, DFGLS=-1.73** (no unit root).
  - GT: N=192, Mean=26.91, St.d.=10.15, Min=2.44, Max=62.88; ADF=-8.30***, DFGLS=-6.69*** (no unit root).
  - RPI per capita (RPIcap): N=192, Mean=44.83, St.d.=3.02, Min=40.01, Max=51.02; ADF= -0.24, DFGLS=2.03 (unit root not rejected → use first difference).
  - First difference of RPIcap (D1.RPIcap): Mean=0.057, St.d.=0.32; ADF=-15.88***, DFGLS=-11.39*** (stationary).
  - RPCE per capita (RPCEcap): ADF=1.10, DFGLS=4.70 (unit root not rejected → use first difference).
  - VIX: N=192, Mean=23.50, St.d.=11.61; ADF=-4.92***, DFGLS=-4.09*** (stationary in levels, but differenced in regressions to capture changes).
  - UR: N=192, Mean=6.07, St.d.=1.93; ADF=0.09, DFGLS=-0.31 (unit root not rejected → use first difference).
- Unit-root tests used: Augmented Dickey–Fuller (ADF) and Dickey–Fuller GLS (DFGLS). Variables with unit roots are entered as first differences in regressions.

### Empirical insights and forecast results
- Structural change: The pandemic altered the relative importance of regressors; models re-estimated including 2020 and 2021 yield different predictor significance compared to pre-pandemic estimates.
- Predictive contributions:
  - RPCE per capita gained predictive importance relative to RPI per capita because pandemic fiscal transfers boosted RPI but did not translate into proportional spending.
  - GT and CAP consistently improve forecast accuracy across samples and model specifications.
  - VIX helps pre-intercept-correction models but is less significant after accommodating the pandemic with a binary indicator and intercept correction.
- Out-of-sample 18-month forecast (point forecast and 5 percent confidence interval):
  - Average point forecast: 92.8 thousand tourist arrivals per month.
  - Lower bound (5 percent confidence level): 78.9 thousand tourists per month.
  - Upper bound (5 percent confidence level): 106.6 thousand tourists per month.

### Policy and operational implications
- Accurate short-term tourism forecasts, incorporating GT and CAP and accounting for structural breaks, are critical inputs for Aruba’s macroeconomic forecasting model ("MARUBA") where tourism inflows are among the most critical inputs for projecting RGDP.
- Forecasting approaches should:
  - Prioritize RPCE per capita over RPI per capita for near-term demand signals during and after pandemic-like shocks.
  - Incorporate real-time indicators such as Google Trends and flight capacity utilization to capture consumer intent and supply-side constraints.
  - Use intercept correction and pandemic indicators to account for structural shifts and normalize residuals.
  - Recognize limitations of some traditional variables (e.g., REER) for short-term forecast improvements in small tourism-dependent economies.

*wpiea2022226-print-pdf - Executive Summary*

### 1.720 for 1, 5, and 10 percent significance levels (denoted with ***, *, ** respectively).

### Modeling and Forecasting Monthly Tourism Arrivals to Aruba Since COVID-19 Pandemic

### Model specifications
- Pure time-series ARMA/ARIMA considered with (p=1, q=1), (p=2, q=1), (p=2, q=2) autoregressive and moving average terms respectively. Equation [1] presents pure time-series ARMA without additional regressors:
  - yt = μ + Σ(γi yt−i, i=1..p) + Σ(θj et−j, j=1..q) + et  [1]
- Multivariate ARIMA (ARIMAX) augments the univariate ARIMA with exogenous variables Xl,t−n. Equation [2]:
  - yt = μ + Σ(γi yt−i, i=1..p) + Σ(θj et−j, j=1..q) + Σ(δl Xl,t−n, l=1..s) + et  [2]
- Seasonal augmentation via monthly binary indicators Dk (k=1...11), dropping September to avoid multicollinearity:
  - ARIMA with seasonal dummies [3]: yt = μ + Σ(γi yt−i) + Σ(θj et−j) + Σ(βk Dk, k=1..11) + et  [3]
  - ARIMAX with seasonal dummies [4]: yt = μ + Σ(γi yt−i) + Σ(θj et−j) + Σ(βk Dk) + Σ(δl Xl,t−n) + et  [4]
- Intercept correction for pandemic via ICovid dummy (value 1 starting 2020m4-2023m12, implicitly assuming 2024 returns to pre-pandemic trend) included in full specification [5]:
  - yt = μ + μIC IC + Σ(γi yt−i) + Σ(θj et−j) + Σ(βk Dk) + Σ(δl Xl,t−n) + et  [5]

### Seasonality and pandemic adjustment
- Two approaches to seasonality discussed:
  - (i) Prior seasonal adjustment or differencing of inputs (required for specifications #1 and #2);
  - (ii) Use of seasonal terms directly in the model (binary monthly indicators preferred here to avoid informational loss).
- Model uses 11 binary monthly indicators D1...D11, dropping September. Expected coefficients for all monthly dummies are positive because tourist arrivals to Aruba reach their lowest interannual levels in September.
- Pandemic intercept correction dummy ICovid: value of 1 from 2020m4-2023m12 (assumption to be re-evaluated by future analysts).

### Exogenous regressors, lag structure, and behavioral assumptions
- Regressors drawn from demand and supply sides, following prior literature:
  - US income/expenditure variables and Google Trends (GT) as main non-autoregressive predictors.
  - Flight statistics from US carriers and flights’ capacity utilization (CAP) introduced.
- Assumed consumer planning timeline and corresponding lags:
  - Consumers start planning about three months in advance: period (T−3) decide budget, (T−2) search and compare, T may cancel.
  - US income/expenditure variables enter models with a third lag (after differencing to remove unit root).
  - Google Trends (GT) lagged by two months (GTt−2).
  - Change in VIX enters without lag (concurrent).
  - US economic policy indices (EPU1, EPU2) or unemployment rate entered with a one-month lag when used.

### Full model(s) and estimation samples
- Full model: ARIMA-X (2, 0, 2) with four exogenous regressors. Main model in equation [6] includes:
  - δ1 d1RPCE t−3 = third lag of first difference in US RPCE (d1RPCE t−3 = RPCE t−3 − RPCE t−4)
  - δ2 GT t−2 = GT lagged by two months
  - δ3 CAP t = concurrent capacity utilization
  - δ4 d1VIX t = first difference in the VIX (d1VIX t = VIX t − VIX t−1)
  - Models include a constant term, two AR terms (yt−1 and yt−2), two MA terms (et−1 and et−2), and 11 seasonal monthly dummies (D1...D11).
  - Equation [6]: yt = μ + Σ(γi yt−i, i=1..2) + Σ(θj et−j, j=1..2) + Σ(βk Dk, k=1..11) + δ1 d1RPCE t−3 + δ2 GT t−2 + δ3 CAP t + δ4 d1VIX t + et  [6]
  - Equation [7] augments [6] with an intercept correction dummy (IC): yt = μ + μIC IC + same remaining terms as [6]  [7]
- Robustness checks:
  - Replace consumption spending with per capita real personal income.
  - Test alternative risk proxies: lagged (one month) differences of US EPU indices and unemployment rate instead of VIX.
- Estimation samples:
  - (i) 2004m1-2019m12;
  - (ii) 2004m1-2020m12;
  - (iii) 2004m1-2021m12.

### Evaluation criteria and model selection
- Statistical and economic significance of estimated coefficients:
  - Statistical significance measured at 1, 5, and 10 percent levels (denoted with ***, **, * respectively).
  - Economic significance requires coefficient signs to agree with a priori assumptions and sizes large enough to matter.
- Residual diagnostics:
  - Low standard deviation (SD), consistency and normality of residuals assessed via Shapiro–Wilk (SWILK) normality test.
- Model parsimony:
  - Minimum Akaike Information Criteria (AIC = −2LL + 2k) used. A decrease in AIC of more than 2 is considered a significant improvement.
- Forecast accuracy:
  - Compare RMSE, MAE, MAPE, and Theil Inequality Coefficient (U-Theil). Lower MAE, RMSE, and U-Theil indicate higher forecast accuracy.
  - Theil U coefficient below one indicates a forecasting model is better than a random guess.
- Assess role of each regressor by auxiliary models excluding exogenous regressors one by one and calculating share of improvement in SD, AIC, RMSE, and Theil U relative to full model versus pure ARMA.

### Forecast assumptions (out-of-sample 18 months: 2022m7-2023m12)
- CAP (concurrent capacity utilization) assumptions:
  - For 2022, CAP is kept at 70 percent (noted as very close to actual value in December 2021); once monthly 2022 values become available, assumed values are replaced by actual observations; for the rest of 2022 and 2023 CAP is assumed at the average known for 2022.
- GT and VIX (or EPU1/EPU2/UR in sensitivity analysis) assumptions:
  - Kept at the latest available level (assuming no change in the forecast period).
- US income (expenditure) per capita assumptions:
  - Assumed to grow at the rate equal to projected growth of real GDP in the USA (according to the latest publicly available “Wells Fargo - U.S. Economic Forecast”) minus expected rate of US population growth (according to US BLS).
  - Paper notes a Wells Fargo vintage from September 30, 2022, assuming personal consumption in the USA (in constant prices) would grow by 0.6, 0.9, and 0.5 percent in 2022q3, 2022q4, and 2023q1, followed by a decline of 1.3, 1.8, and 0.7 percent in 2023q2-2023q4.
  - US population growth figures cited: 0.9 percent in 2022, 0.8 percent in 2023.

### Main empirical results (summary)
- Model fit and specification insights:
  - Coefficient estimates are stable across three ARIMA specifications for all samples.
  - ARIMA (2,0,2) yields the highest likelihood and lowest residuals in all samples. However, when sample includes 2021, ARIMA (1,0,1) produces the lowest AIC; inclusion of AR2 and MA2 terms improves likelihood less than it penalizes AIC in that sample.
  - AR1, AR2, and MA1 terms lose significance when sample includes 2020 and 2021, but MA2 term remains significant.
- Exogenous variables:
  - The US real personal consumption expenditure (RPCE), flight capacity utilization (CAP), and Google search frequencies (GT) are statistically and economically significant in all models.
- Pandemic impact:
  - COVID-19 pandemic reduced total tourism arrivals to Aruba by an average of 43-44 thousand tourists in 2020.
  - COVID-19 pandemic reduced total tourism arrivals by 38-39 thousand visitors over 2020-2021.
  - Inclusion of a COVID-19 dummy leads to a lower marginal effect of higher financial risk/uncertainty as measured by VIX.
  - Before the pandemic, a 1-point increase in maximum monthly VIX would cost Aruba about 5-9 tourists per month. In extended samples and after intercept correction, the coefficient on VIX decreases to -2 and loses statistical significance.
- Flight capacity utilization marginal effect:
  - Average marginal effect of flight capacity utilization increased from 0.37 before the pandemic to [text truncated in source at this point].

*IMF Working Paper — Modeling and Forecasting Monthly Tourism Arrivals to Aruba Since COVID-19 Pandemic (extract).*

### 0.61 and 0.71 in 2020 and 2021 respectively: based on the full sample, one more percent in load factor would

### Modeling and Forecasting Monthly Tourism Arrivals to Aruba Since COVID-19 Pandemic

### Key findings
- Baseline model: ARIMA (2,0,2) with exogenous variables RPCE, CAP, GT, and VIX yields best fit (higher likelihood and lower AIC) and lower forecast error for the full sample.
- Inclusion of RPCE contributes more than 70 percent to the improvement in the model likelihood (LL) and boosts the model’s informational value (AIC) by 14.2 percent, but only 2.1 percent to the residual reduction.
- Flight capacity (CAP) plays the largest role in decreasing the error term and AIC; its marginal impact on arrivals increased over time:
  - CAP coefficient: 0.37*** in 2004m1-2019m12 sample; 0.60–0.61*** in 2004m1-2020m12; 0.71*** in 2004m1-2021m12 sample.
  - Based on the full sample, one more percent in load factor would increase the tourism inflow to Aruba by about 710 persons.
- Google Trends (GT) strengthened as an indicator:
  - L2.GT coefficients: 0.09** / 0.08* / 0.10** / 0.13** across specifications and samples.
  - An increase of internet search frequency by one percent corresponds to 130 additional visitors (based on the full sample).
- VIX impact is moderate and weakened after COVID-19 and intercept correction:
  - D.vixmax coefficients vary (e.g., -0.05, -0.09***, -0.02 across samples/specs).
  - Before the pandemic, a 1-point increase in maximum monthly VIX cost Aruba about 5-9 tourists per month; in extended samples and after intercept correction, coefficient decreases to -2 and loses statistical significance.
- US RPCE vs RPI:
  - RPCE became a better predictor of tourist arrivals during the pandemic; RPI lost predictive usefulness once 2020-21 are included.
  - L3.dRPCE coefficients: 1.72, 1.66, 0.85 (pre-pandemic) and 1.33*, 1.27*, 1.25* (with 2020), and 1.35*, 1.26*, 1.20* (with 2021) across ARIMA 101/201/202 columns.
  - L3.dRPI is statistically significant in pre-pandemic sample (e.g., 2.22*** in a specification) but not in extended samples.
- Per capita consumption spending in the USA (RPCE) gains significance over time:
  - Additional 100 USD per capita real spending in the USA would generate about 125 more tourist arrivals to Aruba (when pandemic months are included).
  - Average per capita spending = 36.3 thousand chained 2012 USD in 2004m1-2021m12; a 1 percent increase (363 USD) would imply about 435-490 more monthly tourist arrivals to Aruba.
- COVID-19 dummy and residuals:
  - Inclusion of a COVID-19 dummy indicates that over 2020m3-2021m12, monthly tourism arrivals decreased by an average of 38-39 thousand visitors over 2020-2021.
  - Residuals increased from 4.0-4.1 before the pandemic to 4.2-2.25 and 4.50-4.54 with extension of the sample to 2020 and 20221, respectively.
  - Pandemic introduced factors the model cannot capture directly (COVID-19 cases, changing preferences, border closures, travel requirements), slightly reducing model likelihood.

### Model performance and statistical measures
- Table 2 (selected stats for models with RPCE, GT, CAP, VIX and intercept correction):
  - SD values: 4.11***, 4.10***, 3.98*** (2004m1-2019m12); 4.25***, 4.23***, 4.20*** (2004m1-2020m12); 4.54***, 4.52***, 4.50*** (2004m1-2021m12).
  - LL: -533.7, -533.2, -528.8 (pre-pandemic); -571.8, -570.5, -569.4 (through 2020); -620.2, -619.3, -618.4 (through 2021).
  - AIC: 1105.4, 1106.4, 1097.7 (pre-pandemic); 1183.6, 1183, 1182.8 (through 2020); 1280.4, 1280.5, 1280.8 (through 2021).
  - SWILK: 0.99** and 0.99 across specifications; residual normality rejected at 10 percent in some pre-pandemic samples for ARIMA 101 and 201.
- Variables’ contribution to model improvement (text-chart summarized):
  - RPCE: >70 percent contribution to LL improvement; 14.2 percent to AIC improvement; 2.1 percent to residual reduction.
  - CAP: largest role in error term and AIC decrease.
  - Intercept Correction (IC): contributes 16.5 percent to error reduction, 2.2 percent to higher LL, and 17 percent to lower AIC.
  - VIX: contributes 5.1 percent to error reduction, 4.3 percent to LL improvement, and 3.9 percent to AIC improvement.
  - Without IC, four exogenous variables reduce residuals by 29 percent, improve LL by 10 percent, and reduce AIC by 11 percent (vs pure ARIMA). After IC, improvements are 34, 13, and 14 percent respectively.

### Forecast accuracy and RMSE impacts
- RMSE comparisons (estimation sample: 2004m1–2021m12) — full sample, training, and test:
  - Full sample ARIMA 202: None = 7.24; RPCE = 7.14; RPCE, CAP = 5.17; RPCE, CAP, GT = 5.08; RPCE, CAP, GT, VIX = 4.99; RPCE, CAP, GT, VIX, IC = 4.69. Change before IC = -2.26; Change after IC = -0.30.
  - Training sample ARIMA 202: None = 6.98; RPCE = 6.85; RPCE, CAP = 5.15; RPCE, CAP, GT = 5.06; RPCE, CAP, GT, VIX = 4.98; RPCE, CAP, GT, VIX, IC = 4.69. Change before IC = -2.00; Change after IC = -0.29.
  - Test sample ARIMA 202: None = 13.77; RPCE = 13.85; RPCE, CAP = 5.66; RPCE, CAP, GT = 5.74; RPCE, CAP, GT, VIX = 5.23; RPCE, CAP, GT, VIX, IC = 7.25. Change before IC = -8.54; Change after IC = 2.02.
- Overall RMSE effects and drivers:
  - A model with RPCE, CAP, GT, and VIX reduces RMSE by about 60 percent in the test sample, with gains around 31 and 29 percent in full and training samples respectively.
  - Intercept correction improves RMSE in training and full samples but can reduce accuracy in the test sample due to binary nature of IC dummy.
  - Flight capacity reduces RMSE by about 80 percent across all samples (single most important factor).
  - Adding GT helps slightly in full and training samples but slightly worsens accuracy in test sample.
  - VIX reduces RMSE by almost 14 percent in the test sample.
  - MA2 term contributes 95, 98, and 97 percent of RMSE reduction for test, training, and full samples respectively.

### Out-of-sample forecasts and uncertainty
- Chosen model (ARIMA 202 with RPCE, CAP, GT, VIX, no intercept correction) produces out-of-sample forecasts for 2022m7-2023m12.
- Point forecast results:
  - Average monthly arrivals over 2022m7-2023m12: 92.8 thousand tourist arrivals per month.
  - Cumulative (baseline scenario): 528.0 thousand tourists in 2022H2; 589.4 thousand in 2023H1; 623.4 thousand in 2023H2.
  - Overall indicated tourism growth of 5.3 percent in 2023.
- Forecast bounds (5 percent confidence level):
  - Average monthly arrival over the 18 months ahead can reach between 78.9 and 106.6 thousand tourists.
- Forecast uncertainty: upper and lower borders obtained as point forecast plus (minus) RMSE times the square root of forecast horizon.

### Robustness and practical recommendations
- Robustness:
  - Models replacing VIX with US unemployment rate, US EPU1, or US EPU2 deteriorate model fit and forecast properties.
  - Models with RPI per capita instead of RPCE lose significance once 2020-21 are included and perform worse in extended samples.
- Practical recommendations for analysts:
  - Use RPCE per capita rather than RPI per capita to capture demand for international vacations during and after the pandemic.
  - Include flight capacity utilization and Google Trends as timely, informative exogenous variables.
  - Include VIX (or similar market volatility measures) to capture uncertainty effects, especially during pandemic periods.
  - Re-estimate models as new observations become available; including more observations generally improves forecast accuracy.
  - Consider intercept correction for pandemic structural break when improving in-sample fit, but assess out-of-sample impacts due to IC’s binary nature.

*Source: IMF Working Paper — Modeling and Forecasting Monthly Tourism Arrivals to Aruba Since COVID-19 Pandemic (selected excerpts).*

### Annex 1. Notes on Google Trends Data

### Annex 1. Notes on Google Trends Data

### Data coverage and frequency
- Google Trends (GT) data are available at a monthly frequency, with the first observation in January 2004.
- Historical GT data start on January 1, 2004, and end 1.5 days before the search.
- The paper did not use real-time data; real-time data are available only for the last week and to perform real-time analysis one would have to scrape data daily.
- The GT data for this paper were first collected during 01/08/2021-02/06/2021, with monthly observations for 2004m1-2021m1; and were extended several times (in March 2022 and August 2022). Upon each extension, the monthly averages are recalculated to include new observations.

### Index construction and normalization
- Google does not report raw query counts. GT provides an index measuring relative popularity of a particular search in a specific region.
- GT index value depends on:
  - (i) the share of the Google queries performed by the users in a particular region on the specific subject in the total number of searches in the same region; and
  - (ii) the number of the search terms in the comparison.
- For a single query, the GT index is normalized to a range [0, 100], with a value of 100 in the period with the highest query share.
- For a query with several comparison items (at most five), the GT index for each term reflects its popularity relative to the most frequently searched term. When downloaded simultaneously for several terms, only one series may observe the index equals 100, with values for the other series being scaled down.

### Example of scaling across comparators
- Example for tourism-related GT search terms for Aruba:
  - When each term is viewed alone, each series would spike at 100 in different periods.
  - When shown together, index values are scaled to the query with the highest interest ("Aruba vacation" in the example).
  - On average, GT series for comparator terms were scaled by factors:
    - "Aruba hotels": 0.75
    - "Aruba beaches": 0.61
    - "Aruba flights": 0.90
    - "Aruba all-inclusive hotels": 0.67

### Search options and categories
- Two modes: search for a term or a topic.
  - Search terms show matches for all terms in the query; topics group terms that share the same concept in any language.
  - This paper used only the search terms.
- Category options: 25 distinct categories (including "Travel," used in this paper), with the 26th option as "All" categories.
- Search types: Web Searches (used here), Image Searches, News Searches, Google Shopping, or YouTube Searches.

### Sampling, exclusions, and averaging
- GT compilation employs sampling techniques.
- Two features to account for:
  - (i) series exclude searches made very few times, repeated by the same user over a short period, and queries with special characters; and
  - (ii) values can differ from day to day and even within the day.
- Researchers often average the GT index obtained during about four weeks to smooth sampling variability.
- The paper’s approach of recalculating monthly averages upon extension of the sample is supported by Eichenauer et al., 2022.
- Reference for GT details: “FAQ about Google Trends data” at https://support.google.com/trends/answer/4365533?hl=en.

*Source: wpiea2022226-print-pdf - Annex 1. Notes on Google Trends Data*

### Annex 5. Robustness Check

### Annex 5. Robustness Check

### Table A5-1. Tourist Arrivals to Aruba — Model w/RPI, GT, CAP, VIX, IC and Monthly Dummies (samples: 2004m1-2019m12; 2004m1-2020m12; 2004m1-2021m12)
- Seasonal monthly dummies (i1–i12):
  - i1: 13.48***; 13.44***; 13.47*** — 14.65***; 14.54***; 14.47*** — 15.16***; 15.15***; 15.05***
  - i2: 9.30***; 9.25***; 9.56*** — 10.52***; 10.38***; 10.40*** — 10.88***; 10.86***; 10.81***
  - i3: 16.23***; 16.21***; 16.29*** — 17.78***; 17.74***; 17.77*** — 17.61***; 17.59***; 17.62***
  - i4: 13.58***; 13.62***; 13.63*** — 13.26***; 13.28***; 13.26*** — 13.62***; 13.66***; 13.64***
  - i5: 2.49; 2.48; 2.59 — 2.19; 2.17; 2.17 — 3.01; 3.03; 3.02
  - i6: 5.10***; 5.08***; 5.31*** — 4.29**; 4.18**; 4.27** — 5.22***; 5.21***; 5.26***
  - i7: 15.81***; 15.79***; 15.91*** — 13.94***; 13.80***; 13.81*** — 14.29***; 14.24***; 14.23***
  - i8: 14.24***; 14.19***; 14.59*** — 13.37***; 13.30***; 13.33*** — 13.03***; 13.00***; 13.01***
  - i10: 3.62**; 3.59**; 3.88** — 2.60*; 2.53; 2.55* — 2.30*; 2.25; 2.24
  - i11: 9.23***; 9.20***; 9.27*** — 9.72***; 9.64***; 9.58*** — 10.14***; 10.13***; 10.06***
  - i12: 21.58***; 21.54***; 21.74*** — 22.50***; 22.43***; 22.35*** — 22.85***; 22.85***; 22.78***

- RPI (L3.dRPI):
  - 2004m1-2019m12: 2.68***; 2.77***; 2.22***
  - 2004m1-2020m12: -0.07; 0.03; 0.03
  - 2004m1-2021m12: -0.13; -0.14; -0.17

- CAP:
  - 2004m1-2019m12: 0.37***; 0.37***; 0.37***
  - 2004m1-2020m12: 0.59***; 0.60***; 0.59***
  - 2004m1-2021m12: 0.70***; 0.71***; 0.71***

- GT (L2.GT):
  - 2004m1-2019m12: 0.09*; 0.08**; 0.09**
  - 2004m1-2020m12: 0.10**; 0.10**; 0.09**
  - 2004m1-2021m12: 0.12**; 0.13**; 0.12**

- VIX (D.vixmax):
  - Coefficients across samples: -0.03; -0.03; -0.06* — -0.02; -0.02; -0.02 — -0.03; -0.03; -0.03

- Intercept Correction (IC):
  - 2004m1-2020m12: -43.90***; -43.88***; -43.52***
  - 2004m1-2021m12: -38.90***; -39.15***; -38.85***

- Constants:
  - 2004m1-2019m12: 34.50***; 34.32**; 34.67***
  - 2004m1-2020m12: 18.17; 17.51; 18.23
  - 2004m1-2021m12: 14.50; 15.41; 15.3

- ARIMA dynamics (examples across ARIMA specifications 101, 201, 202):
  - L1.ar: values include 0.97***; 1.29***; 0.10* — 0.97***; 1.41***; 0.47 — 0.98***; 1.39***; 0.59
  - L2.ar: present as X or values e.g. -0.31; -0.42*; -0.40*
  - L1.ma: -0.31**; -0.60**; 0.62*** — -0.31**; -0.71***; 0.24 — -0.20; -0.61***; 0.22
  - L2.ma: present in some specifications (e.g., -0.38***; -0.31**; -0.24**)

- Model fit and diagnostics (by sample and ARIMA specification):
  - SD: 3.99***; 3.97***; 3.88*** — 4.31***; 4.27***; 4.25*** — 4.60***; 4.56***; 4.53***
  - LL: -528.1; -527.2; -524.2 — -574.4; -572.8; -571.6 — -622.9; -621.4; -620.1
  - AIC: 1094.2; 1094.3; 1088.4 — 1188.8; 1187.6; 1187.2 — 1285.8; 1284.9; 1284.2
  - SWILK: 0.99**; 0.99; 0.99 — 0.99; 0.99; 0.99 — 0.99; 0.99; 0.99

### Table A5-2. Tourist Arrivals to Aruba — Models with alternative risk measures (2004m1-2021m12)
- Three alternative exogenous sets reported: RPCE, CAP, GT, IC, & UR; RPCE, CAP, GT, IC, & EPU1; RPCE, CAP, GT, IC, & EPU2. Each reported for ARIMA 101, 201, 202.

- Seasonal monthly dummies (i1–i12) — selected entries:
  - i1: 15.31***; 15.29***; 15.23*** — 15.25***; 15.22***; 15.13*** — 15.25***; 15.22***; 15.14***
  - i2: 10.93***; 10.89***; 10.85*** — 11.03***; 11.00***; 10.99*** — 11.03***; 11.00***; 10.98***
  - i3: 17.70***; 17.66***; 17.67*** — 17.68***; 17.65***; 17.67*** — 17.68***; 17.65***; 17.67***
  - i12: 23.03***; 23.02***; 22.96*** — 23.05***; 23.04***; 22.96*** — 23.05***; 23.04***; 22.96***

- RPCE (L3.dRPCE):
  - RPCE, CAP, GT, IC, & UR: 1.39*; 1.31*; 1.24*
  - RPCE, CAP, GT, IC, & EPU1: 1.44*; 1.36*; 1.33*
  - RPCE, CAP, GT, IC, & EPU2: 1.45*; 1.37*; 1.34*

- CAP:
  - 0.71***; 0.72***; 0.73*** — 0.68***; 0.69***; 0.68*** — 0.68***; 0.69***; 0.69***

- GT (L2.GT):
  - 0.13**; 0.13**; 0.12** — 0.13**; 0.13**; 0.12** — 0.13**; 0.13**; 0.12**

- Unemployment dummy (D.UR) included only in the RPCE,...,& UR specification:
  - D.UR: 0.20; 0.22; 0.24 (other specifications show X for D.UR)

- Economic policy uncertainty dummies:
  - D.EPU1: included only in RPCE,...,& EPU1 spec: 0.00; 0.00; 0.01
  - D.EPU2: included only in RPCE,...,& EPU2 spec: 0.00; 0.00; 0.00

- Intercept Correction (IC):
  - IC: -38.97***; -38.93***; -38.32*** — -39.94***; -40.17***; -39.96*** — -39.94***; -40.16***; -39.91***

- Constants:
  - 13.99; 14.42; 13.9 — 16.45; 17.33; 17.38 — 16.45; 17.3; 17.29

- ARIMA dynamics (examples):
  - L1.ar: 0.98***; 1.30***; 0.62* — 0.98***; 1.31***; 0.59* — 0.98***; 1.31***; 0.59*
  - L2.ar: X; -0.31*; 0.36 — X; -0.32; 0.39 — X; -0.32; 0.39
  - L1.ma: -0.24*; -0.54***; 0.16 — -0.23*; -0.54***; 0.2 — -0.23*; -0.54***; 0.20
  - L2.ma: present in ARIMA 202 specifications as -0.22**

- Model fit and diagnostics (all three exogenous sets, ARIMA 101/201/202):
  - SD: 4.54***; 4.52***; 4.50***
  - LL: -620.4; -619.4; -618.5 (similar LL reported across the three alternative risk-measure specifications)
  - AIC: 1280.8; 1280.9; 1280.9 (AIC reported as 1280.9 across alternative EPU specifications)
  - SWILK: 0.99**; 0.99**; 0.99**

### Key statistics and robustness signals from Annex 5 tables
- Seasonal pattern is strong and highly significant across months i1–i12 with multiple coefficients significant at *** and ** levels.
- CAP shows consistent positive significance across specifications and samples (e.g., 0.37***; 0.59***; 0.70*** in Table A5-1; 0.71*** etc. in Table A5-2).
- Search-intensity measure GT (lagged) shows consistent positive significance across samples (e.g., L2.GT coefficients around 0.08–0.13 with * or ** significance).
- Intercept Corrections (IC) are large negative and highly significant across specifications (e.g., -43.90***; -38.97***; -39.94***).
- Risk and uncertainty proxies:
  - VIX (D.vixmax) coefficients are small and mostly not significant, with occasional marginal significance (e.g., -0.06*).
  - RPCE (L3.dRPCE) shows positive and sometimes significant coefficients in Table A5-2 (e.g., 1.39* to 1.45*), but RPI (L3.dRPI) in Table A5-1 shows changing sign and significance across sample periods.
  - EPU dummies (D.EPU1, D.EPU2) are reported with near-zero coefficients where included.
- ARIMA model selection differences affect short-run dynamics (L1.ar, L2.ar, L1.ma, L2.ma coefficients vary by ARIMA specification), but SD, LL, AIC and SWILK indicate broadly similar fit within each panel of specifications.

*Source: IMF Working Paper — Annex 5. Robustness Check, Modeling and Forecasting Monthly Tourism Arrivals to Aruba Since COVID-19 Pandemic*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2022/english/wpiea2022226-print-pdf.pdf_