View on GitHub View Portfolio

Hurricane Analysis

A Bayesian analysis of hurricane data from the National Hurricane Center.

Tyler Pollard 2024-08-25

Introduction

The data for this analysis are from A Feed Forward Neural Network Based on Model Output Statistics for Short-Term Hurricane Intensity Prediction, with detailed variable descriptions available here. This paper uses deep learning to improve 24-hour ahead forecasts of hurricane intensity (maximum wind velocity, VMAX).

The primary prediction model, HWRF, which is a mathematical weather prediction model based on differential equations. In addition to the forecast, HWRF has many other state variables such as sea surface temperature, longitude, time of year, etc, that are usually discarded.

This analysis aims to assess whether incorporating these HWRF state variables can enhance the accuracy of hurricane intensity predictions

Data Exploration (Shiny App)

To facilitate exploratory analysis, I developed an interactive Hurricane Analysis app. This tool enables the visualization of key variables and their relationships with VMAX:

Model Description

The response variable $Y_i$ is the observed VMAX for each hurricane observation. The dataset consists of $i = 1, …, n = 1705$ total observations from 87 unique storms, where each storm is identified by a StormID ($S_i$). Every observation is associated with $p = 20$ covariates ($X_{ij}$, where $j = 1, …, p$) that capture key spatial characteristics, environmental conditions, and forecast-based features relevant to hurricane intensity.

Observations are recorded at 6-hour intervals, representing StormElapsedTime. However, some storms have missing increments, making the time sequence irregular. StormElapsedTime was not included as a predictor in the model because the HWRF model already accounts for time evolution through its differential equation framework.

To ensure comparability across variables, all covariates underwent the following preprocessing steps:

  1. Arcsinh transformation was applied to handle skewed distributions.
  2. Centering and scaling were performed to standardize the variables.

Given the positive right-skewed distribution of VMAX, two likelihood functions were considered:

Log-normal

\[\begin{aligned} Y_{i} &\sim \text{LogNormal}(\mu_i, \sigma^2) \\ \mu_i &= \beta_{0} + \sum_{j=1}^{p}X_{ij}\beta_{j} + \theta_{S_i} \\ \theta_{S_i} &\sim \text{Normal}(0, \tau^2) \end{aligned}\]

Gamma

\[\begin{aligned} Y_{i} &\sim \text{Gamma}(\alpha, \alpha/\mu_i) \\ log(\mu_i) &= \beta_{0} + \sum_{j=1}^{p}X_{ij}\beta_{j} + \theta_{S_i} \\ \theta_{S_i} &\sim \text{Normal}(0, \tau^2) \end{aligned}\]

where $\beta_{j}$ is the effect of covariate $j$ with weakly informative priors $\beta_j \sim \text{Normal}(0,5)$. We set a weakly prior on $\alpha \sim \text{InvGamma}(0.1, 0.1)$. In the random effects models, we set weakly priors $\tau \sim \text{InvGamma}(0.1, 0.1)$, otherwise $\theta_{S_i} = 0$.

Model Comparisons

All four models were fit to the data and assessed for convergence to ensure reliable parameter estimates. After confirming convergence, model selection was performed using the Widely Applicable Information Criterion (WAIC) and Expected Log Predictive Density (ELPD), both of which estimates out-of-sample predictive accuracy. Table 1 presents the WAIC values along with associated parameters for each model:

Table 1: Model Selection Criteria Using WAIC and ELPD
Model waic1 p_waic2 elpd_waic3 elpd_diff4 se_diff5
Log-normal (Random Effects) 12697.84 91.02 −6348.92 0.00 0.00
Gamma (Random Effects) 12712.67 92.91 −6356.33 −7.41 4.18
Log-normal (Fixed Effects) 12835.91 22.22 −6417.95 −69.03 12.87
Gamma (Fixed Effects) 12863.94 22.85 −6431.97 −83.05 13.85
1 WAIC (waic): Widely Applicable Information Criterion, a model selection metric balancing fit and complexity. Lower values indicate better expected predictive accuracy.
2 Effective Parameters (p_waic): An estimate of the number of effective parameters in the model; higher values indicate more flexibility.
3 ELPD (elpd_waic): Expected log predictive density, quantifying out-of-sample predictive performance. Higher (less negative) values indicate better predictive accuracy.
4 ELPD Difference (elpd_diff): The difference in elpd_waic relative to the best model (logNormal_Rand_Fit). The best model always has elpd_diff = 0.
5 SE of Difference (se_diff): The standard error of elpd_diff, measuring uncertainty in the difference estimates. Large absolute elpd_diff values relative to se_diff indicate meaningful performance differences.

Based on WAIC and elpd_diff, the Log-normal model with random effects demonstrated the best overall predictive performance. This decision was supported by the following observations:

Model Refinement

After selecting the Log-normal model with random effects, additional refinements were made to improve model interpretability and predictive performance. The refinement process followed two key steps:

  1. Addressing Multicollinearity

    • Covariates with high variance inflation factors (VIF) were identified and removed to reduce multicollinearity, ensuring stable parameter estimates.
    • TCOND7002, INST2, and SHTFL2 exhibited high VIF values and were removed before refitting the model.
  2. Iterative Variable Selection

    • To improve model parsimony, variables were removed one at a time, with the model iteratively refit after each removal.
    • This process continued until the 95% credible intervals of all remaining covariates no longer contained 0, ensuring that only statistically meaningful predictors were retained.

Final Model

After refinement, the final model retained $p = 9$ covariates:

This final model strikes a balance between parsimony and predictive power, ensuring that only the most relevant predictors are retained while minimizing unnecessary complexity. By including variables that significantly contribute to explaining variations in VMAX, the refined model improves both interpretability and generalizability for forecasting hurricane intensity.

Goodness of Fit

To assess the model’s goodness of fit, posterior predictive checks (PPCs) were performed by drawing samples from the posterior predictive distribution (PPD). These checks compare key summary statistics of the observed data to those generated from the model, ensuring that the fitted model can replicate important characteristics of the data.

The figure below displays the empirical distribution of observed VMAX alongside the posterior predictive distribution, as well as Bayesian p-values for the mean, standard deviation (SD), and range:

<strong>Figure 1: Posterior Predictive Checks for Log-normal Random Effects Model</strong>

Figure 1: Posterior Predictive Checks for Log-normal Random Effects Model

These p-values fall within an acceptable range, staying sufficiently away from 0 and 1, indicating that the model does not systematically overestimate or underestimate variability in the data. Additionally, the empirical distribution of observed VMAX aligns well with the simulated distributions from the PPD draws, further supporting model adequacy.

Overall, the PPCs confirm that the model provides a reasonable fit to the observed data and successfully captures key characteristics of VMAX distribution.

Variable Importance

After fitting the final model with random intercepts and weakly informative priors, the importance of each covariate was examined. Table 2 presents the posterior mean, standard deviation, and 95% credible interval for each parameter after partially pooling over StormIDs.

Since all covariates were centered and scaled, the posterior means allow for direct comparison of covariate importance, where larger absolute values indicate stronger effects on VMAX.

From Table 2, the most influential covariates in modeling VMAX were Land, HWFI, and HWRF. Given that HWRF is already a widely used forecast for VMAX, its significance in the model is expected. However, an interesting result is that HWFI appears to have a larger effect on VMAX than HWRF, suggesting it may contribute more predictive value in this model.

Table 2: Posterior summary for model parameters
Parameter Mean SD Q2.5 Q97.5
Intercept 4.0114 0.0111 3.9891 4.0330
LandLand -0.1175 0.0243 -0.1655 -0.0699
SHR_MAG -0.0456 0.0058 -0.0569 -0.0340
RHLO 0.0275 0.0066 0.0146 0.0407
CAPE3 0.0270 0.0070 0.0133 0.0404
CP1 0.0141 0.0072 0.0001 0.0282
TCONDSYM2 0.0305 0.0076 0.0157 0.0452
COUPLSYM3 0.0237 0.0107 0.0030 0.0447
HWFI 0.2204 0.0127 0.1954 0.2450
HWRF 0.1849 0.0132 0.1597 0.2105

Prediction

In the dataset provided for this project, there were an additional 668 observations from new StormIDs where VMAX was missing. The posterior predictive mean and 95% credible interval were used to estimate VMAX for these out-of-sample (OOS) observations. The predictive performance of the model was assessed using cross-validation and, later, by evaluating its predictions against the actual OOS values.

Cross-Validation

To estimate predictive accuracy, 5-fold cross-validation (CV) was performed by splitting the 1,705 observations into 5 approximately equal folds based on randomly selected StormIDs. This mimics the real-world scenario of predicting entirely new hurricanes.

The model’s performance was evaluated using:

Alongside CV, predictions were also generated using the PPD by treating the fit observations as missing. Table 3 presents the prediction metrics for both methods:

The 95% credible interval coverage from CV (0.949) aligns closely with the expected 0.95, confirming that the model reliably quantifies uncertainty in its predictions. Additionally, the MAE from CV (8.778) represents a 12% improvement over the HWRF baseline and demonstrates the model’s ability to generalize to new storms.

Table 3: Cross-Validation prediction metrics on observed data
Method HWRF MAE Model MAE COV
Model PPD 10.006 8.017 0.951
5-fold CV 10.006 8.778 0.949

Out-of-Sample

Following the initial analysis, the actual VMAX values for the 668 out-of-sample observations were obtained. This allowed for a true performance evaluation of the model, rather than relying solely on cross-validation estimates.

The PPD mean and 95% credible interval were computed for these OOS observations, and prediction accuracy was assessed using MAE, Mean Absolute Deviation (MAD), and COV. The results are presented in Table 4:

Table 4: Prediction metrics on out-of-sample data
Method HWRF MAE Model MAE Model MAD COV
Model PPD 7.530 7.131 6.966 0.970

The model outperforms HWRF (7.131 vs. 7.53 MAE), confirming its predictive strength. The coverage (0.97) suggests that the model’s uncertainty quantification remains reliable even when predicting unseen storms.

Prediction Plots

To visualize the model’s performance on OOS predictions, we compare the PPD mean and 95% credible intervals against the actual observed VMAX values.

Figure 2 provides a comprehensive view of all OOS storms, showing how well the model’s predictive mean aligns with actual VMAX values.

To complement this, the Figure 3 focuses on longer-duration storms, offering a closer look at how predictions evolve over time. This plot also includes HWRF forecasts alongside actual VMAX values, allowing for a direct comparison between the model and HWRF predictions.

<strong>Figure 2: Out-of-Sample Forecasts - Full Dataset.</strong> The model’s posterior predictive mean (green) tracks closely with observed VMAX values (gray), with the 95% credible interval (blue) capturing most data points, indicating well-calibrated uncertainty estimates.

Figure 2: Out-of-Sample Forecasts - Full Dataset. The model’s posterior predictive mean (green) tracks closely with observed VMAX values (gray), with the 95% credible interval (blue) capturing most data points, indicating well-calibrated uncertainty estimates.

<strong>Figure 3: Out-of-Sample Forecasts - Long Storms.</strong> The model’s posterior predictive mean (green) is compared against both HWRF forecasts (red) and actual VMAX values (gray). The model successfully tracks storm intensity while providing well-calibrated uncertainty estimates, demonstrating improvement over HWRF in many cases.

Figure 3: Out-of-Sample Forecasts - Long Storms. The model’s posterior predictive mean (green) is compared against both HWRF forecasts (red) and actual VMAX values (gray). The model successfully tracks storm intensity while providing well-calibrated uncertainty estimates, demonstrating improvement over HWRF in many cases.

Key Takeaways

These results further support the statistical improvements in MAE reported in Table 4, reinforcing the model’s reliability for predicting hurricane intensity.

Discussion

This analysis explored whether incorporating additional HWRF state variables improves hurricane intensity predictions beyond the existing HWRF model alone. A Log-normal random effects model was selected as the final model based on WAIC and ELPD criteria, demonstrating the best predictive performance.

Through cross-validation and out-of-sample evaluation, the model consistently outperformed HWRF forecasts:

The prediction plots further validated the model’s reliability, demonstrating well-calibrated posterior intervals and a strong alignment between predicted and observed VMAX values, particularly for long-duration storms, where forecasting is most critical.

Limitations and Future Work While the model performed well, there are areas for potential improvement:

Limitations

While the model performed well, there are areas where future improvements may be warranted:

Future Work

Several avenues could further improve the model’s performance and practical application:

This study demonstrates that leveraging additional state variables from HWRF significantly enhances hurricane intensity prediction, offering a data-driven approach to improving forecasting accuracy. Future refinements can build upon these results to further enhance predictive performance in operational forecasting settings.