2013 Hurricane Season Predictions Brandon Sherman Vinicius Taguchi Kehao Zhu http://geology.com/hurricanes/hurricane-names.shtml
Purpose Analysis Methods Model Selection Results Conclusion
Purpose To accurately predict the number of hurricanes that will impact the United States during the 2013 hurricane season Days Before Landfall Error in miles 5 350 4 290 3 230 2 160 1 100 http://web.mit.edu/12.000/www/m2010/teams/neworleans1/predicting%20hurricanes.htm http://designintentions.wordpress.com/2010/10/07/case-study-hurricane-katrina-and-post-disaster-social-learning/ Damage from Hurricane Katrina in 2005 NOAA Data, 2004
Methods: Initial Model Averaged 12 months covariates from January of previous year to December of previous year e.g. Jan. 2011-Dec. 2011 to predict 2012 hurricane season Started with all the covariates in the poisson regression model Ignore any covariates which have NA values
Methods: Model Selection BIC backward/forward(both) selections AIC backward/forward(both) selections Physical Model
Methods: Physical Model MDRSST - MDR Sea Surface Temperature MDRSLP - MDR Sea Level Pressure MDRVWS - MDR Vertical Wind Shear NINO4,NINO12 - El Niño/La Niña AMO - Atlantic Multi-Decadal Oscillation QBO - Quasi-Biennial Oscillation AMM - Atlantic Meridional Mode TSA,TNA,DM,GGST,NGST,SGST - Global/Hemispheric Surface Temperatures
Methods: Cross-Validation 5-fold cross-validation for: null model full model "BIC"-selected model "AIC"-selected model physical model
Results: Statistics
Results: Model Selection Number of covariates AIC Error rates from 5-folds cross-validation Null model 316.84 4.181301 Full model 30 311.59 5.654012 "BIC" selected model 6 307.98 4.106643 "AIC" selected model 14 285.73 4.318777 Physical Model 13 301.51 4.358287
Conclusions BIC model most preferred Most parsimonious Lowest cross-validation error Atlantic Dipole Mode is the difference between TSA and TNA, so it appears the Atlantic Dipole Mode itself is sufficient for the model
Final Model Prediction for 2013: 7.458387 Landfalls Coefficients Estimates Intercept 1.28687 Atlantic Multi-Decadal Oscillation 1.32681 Atlantic Dipole Mode -0.17114 El Niño - Region 4 0.27248 Quasi-Biennial Oscillation -0.02461 Prediction for 2013: 7.458387 Landfalls
Discussion Strength: Weakness: relative simple model, easy to interpret Weakness: the method of averaging the covariates is arbitrary
Thank You For Listening
Works Cited Post-Katrina. N.d. Photograph. Design Intentions. Web. 16 May 2013. Samost, Aubrey. "Predicting Hurricanes: A Not So Exact Science." Predicting Hurricanes. Massachusetts Institute of Technology, n.d. Web. 16 May 2013. Satellite Image of Hurricane Fran. N.d. Photograph. NASA. Geology.com. Web. 16 May 2013.