M odelling This article appeared in the East Asia Special 2001 edition of Global Reinsurance. Improving typhoon predictions

Mark Saunders and Paul Rockett describe techniques to better estimate future storm activity.

ropical cyclones are the most costly cause the vast majority of the damage and industry and, understandably, confidence and deadly type of natural disaster death from tropical cyclones in east Asia. in their quality needs building before Taffecting much of Japan, South The 1991-2000 period has seen an annual insurance and reinsurance executives Korea, Taiwan, the and coastal average of 9.1 intense typhoons and 17.8 employ them routinely in business areas in other east Asian countries. The typhoons in the northwest Pacific basin, and decisions. The purpose of this article is to annual damage bill and fatality rate impacts an annual average typhoon strike rate on describe the typhoon predictions now in east Asia 1990-1999 averages respectively Japan, Taiwan and the Philippines of 2.4, available, their methodologies and skill, US$3.3bn and 740 deaths. However, 1.7 and 1.9 respectively. Already the 2001 and the prospects for further substantial inter-annual variability exists in season has seen several typhoons make improvements in the coming years. regional losses. For landfall on east Asian shores (figure 1). 2001 has been a busy year for tropical example, in 1991 and 1988 Japan Skilful typhoon predictions will benefit storm landfalls in east Asia. This GMS experienced losses of $7.5bn and zero insurers and reinsurers, as well as society, satellite image (Figure 1) shows typhoon respectively, measured in 2000 dollars. government and other businesses by Utor striking and southern Typhoons, described as one minute of reducing the risk, uncertainty and financial on 5 July, 2001. Utor killed at least sustained winds of over 73mph (63 knots), volatility inherent to varying active and 23 in China, 121 in the Philippines and and intense typhoons one minute of inactive tropical storm seasons. Predictions one person in Taiwan. It was one of the sustained winds of over 110mph (95 knots), are a relatively new resource for the strongest storms to strike the region in

www.globalreinsurance.com statistical analysis of the past behaviour of probabilistic in nature. They are derived storms (statistical forecasts) or from an from a statistical assessment of the major application of our knowledge of the climatic factors affecting tropical cyclone physics of the atmosphere and ocean activity in each region. These factors may (dynamical forecasts). be either contemporaneous or lagged. Lagged factors are climatic conditions Seasonal forecasts prevailing at the time of the forecast that Seasonal forecasts provide an idea of the are known to be precursors of typhoon level of typhoon activity to anticipate activity in the coming season. during the coming season. In the Contemporaneous factors are climatic northwest Pacific, typhoons can occur in conditions that prevail during the main all months of the year, but historically 90% typhoon season itself. In order for of activity happens after 1 June each year, contemporaneous factors to be useful in making this, for practical purposes, the seasonal typhoon forecasts they must be start of the main typhoon season. sufficiently predictable themselves at the A number of organisations now issue necessary lead times. ENSO (El Niño- seasonal forecasts for the Atlantic, Southern Oscillation) is an example of a northwest Pacific and Australian basin contemporaneous predictor useful in regions. Northwest Pacific seasonal typhoon seasonal forecasts. High (low) typhoon forecasts are issued by the typhoon activity is observed when ENSO University of Hong Kong (see is in its positive (negative) phase during the http://aposf02.cityu.edu.hk/~mcg/index.ht main typhoon season from July to m) and by the TropicalStormRisk.com October. (TSR) venture (see http://www.tropicalstormrisk.com). These Intensity forecasts forecasts began routinely in 2000. The Accurate typhoon forecasts not only save University of Hong Kong predictions are lives, but can be used as a risk issued in early April, and are updated in quantification tool for insurers and June. The TSR forecasts are issued in late reinsurers once a storm has formed. If the May/early June, and for the 2002 season risk of landfall can be quantified in near will be available updated monthly from real-time, enterprising executives can use early April to early August. These forecasts loss estimates to optimise capital, organise provide an indication of the numbers of claims response units and even trade tropical storms of various intensities that catastrophe bonds. can be expected in the season ahead. When a tropical cyclone is active, While it is impossible to give forecasts of forecasts of the storm position and precise positions of landfall at such long maximum winds at 12, 24, 36, 48, 72, 96 range, predictions can be made of the risk and even 120 hours ahead are issued every to broad regions, such as Japan or the south six hours by the meteorological agency China coast. These forecasts are often with warning responsibility in the region of Figure 1: Typhoon Utor striking Hong Kong

recent years with flood and wind damage in the southern Chinese province of alone estimated at US$290m. Utor also caused most of Hong Kong to shut down for business for two days. (Image courtesy of Tim Olander, University of Wisconsin-Madison Cooperative Institute for Meteorological Satellite Studies, US.) Methodologies Typhoon predictions come in two types: seasonal forecasts of overall activity for the coming season, and track and intensity forecasts for individual events with up to five days lead time. Forecast models are of Source: authors’ own two main forms, based either on a

www.globalreinsurance.com “ The only near certainty in forecasting is that the forecast will be wrong.” the storm. For the northwest Pacific, the two main sources are the Japan Meteorological Agency (http://ddb.kishou.go.jp/typhoon/cyclone/c yclone.html) and the US Navy and Army Joint Typhoon Warning Center (JTWC) on Guam the atmosphere by a set of mathematical a result of the well-known ‘Butterfly (http://www.npmoc.navy.mil/jtwc.html). equations, the parameters of which are Effect’. Attempts have been made to use Forecasters at these centres consider calculated. dynamical models to investigate future independent predictions made by a Dynamical models can be applied to the changes in typhoon activity due to global number of dynamical and statistical whole atmosphere (global models) or to warming, but to date the results have prediction models produced both in-house smaller local areas (local models). Despite proved inconclusive. and by other organisations. The JTWC, recent advances in model resolution for example, takes guidance from the UK (current gridded global dynamical models Statistical models Meteorological Office and the NOAA typically have a resolution of about one Statistical methods are used in all types of Geophysical Fluid Dynamics laboratory degree in latitude and longitude with typhoon predictions: seasonal, track and amongst others. around 20 levels through the atmosphere), intensity forecasts. In all cases, analyses of global models are still too coarse to past data are made to quantify the Dynamical models represent the important physical processes probability of an event occurring, given Dynamical models use state-of-the-art within a typhoon. In some global models certain conditions. Linear regression computing techniques to apply our this is overcome by using a ‘bogussing’ analysis is applied to past data to identify knowledge of the physics of the scheme. Under such a scheme, once a links between the desired forecast quantity atmosphere and ocean to forecast what suspect tropical cyclone is identified and potential predictors. Predictors which may occur ahead, knowing its initial state. artificial data is introduced into the model have a strong statistical link to the forecast To achieve this, the atmospheric system to better represent the storm. quantity are incorporated into a linear must be simplified sufficiently to allow the Dynamical models have good uses in prediction model. In the case of seasonal model runs to be completed in time to many operational track forecasting forecasting, the predictors are climatic make a useful forecast. This is done either situations. They are not employed factors which strongly influence tropical by calculating the atmospheric variables generally for seasonal predictions, since cyclone frequency in the regions of (temperature, humidity, windspeed, etc.) at even tiny errors in the initial state of the interest. certain grid points only, or by representing model can become massively amplified The advantage of statistical methods is the distribution of these variables through when forecasting at long lead times. This is that they are easier and quicker to run than dynamical models. In the case of seasonal predictions, they are more useful than Figure 2: typhoon seasonal prediction skill dynamical models because they are free of the large amount of noise that results from the Butterfly Effect inherent in these models. On the downside, since they are based on (often limited) past data, they are likely to have little skill in situations not previously encountered (under changed climate conditions, for example). Dynamical models are not susceptible to this problem. Dynamical track and intensity models typically outperform their statistical equivalents. Figure 2 illustrates the TSR (TropicalStormRisk.com) seasonal forecast skill for the annual number of typhoons forming in the northwest Pacific. The plot displays the mean skill (and 95% confidence interval) as a function of monthly lead from August to March. The skill measure is the percentage Source: authors’ own improvement in root mean square error

www.globalreinsurance.com Figure 3: NorthWest Pacific tropical cyclone forecast its full potential has not been explored. We positional errors – Met Office global model can be certain that skilful seasonal forecasts of typhoon and intense typhoon activity are available from the beginning of May (figure 2), and that skill increases as the season progresses. The products and services on offer from TSR and the University of Hong Kong will continue to expand. Monthly updated forecasts will be available for the 2002 season, as will forecasts for new landfalling regions. The planned incorporation of dynamical model data into the TSR forecasts will be a further novel development, and may bring useful benefits. The key factor which would provide seasonal typhoon skill at leads prior to May is improved prediction of ENSO. Typhoon track errors have not reached Source: authors’ own predictability limits, and will continue to fall as computers become more powerful, as satellites provide better real-time over the 1971-2000 climatology forecast. Overall, the 2000 northwest Pacific monitoring of tropical environmental Positive skill indicates the forecast model typhoon season saw slightly below average conditions, and as computer models performs better than climatology, while activity. The University of Hong Kong and improve. Typhoon intensity forecasts are negative skill indicates it does worse than TSR seasonal forecasts are calling for near- also expected to become more accurate in climatology. All hindcasts are made solely average typhoon activity in 2001. future seasons. using prior data. The figure shows that Figure 3 shows forecast positional errors With the recognition that a major cause reliable seasonal typhoon forecasts may be for tropical cyclones in the northwest of death, injury and damage in typhoon made as early as the beginning of May; Pacific 1988-2000 from the UK Met landfalls is freshwater flooding, a number of 90% of annual typhoon activity occurs Office global forecast model (see research groups are beginning to focus on after 1 June. http://www.metoffice.govt.uk/sec2/sec2cycl the prediction of rainfall in tropical cyclones one/tcerrors/ for further information). at landfall, and advances may be expected. Prediction skills Mean annual errors are shown for forecasts Both seasonal and track forecasts of The only true proof of a forecast is its at leads of 0, 24, 48, 72, 96 and 120 typhoons have skill which can benefit long-term track record, but how can one hours. The mean is for all storms which insurers and reinsurers in their business judge the potential accuracy and usefulness formed during the year. The figure shows decisions and assessment of risk. By of the University of Hong Kong and TSR that forecast track errors are decreasing sharing forecasts with customers, risk seasonal forecast models which have been steadily. Errors at all leads have fallen by awareness increases, and risk mitigation operational only since early 2000? One around 50% over the last decade. (Figure strategies can be developed. The skill of all good result does not mean that the method courtesy of Julian Heming, Met Office, forecasts will increase in the coming years is foolproof, any more than one bad result UK). with further research. means that it is useless. The only near Typhoon track forecasts have been certainty in forecasting is that the forecast operational for over a decade, so an By Mark Saunders and Paul Rockett will be wrong. Even forecasts with limited assessment of their skill is straightforward. skill can have real value. If a forecast is Figure 3 shows the performance of a well- Mark Saunders (top, better than climatology (that is than long- respected track model – the UK Met Office right) is a Senior term averages) only six years out of ten, an global model – for tropical cyclone Lecturer in the investor in a portfolio of weather positional errors in the northwest Pacific. Department of Space derivatives should make money. The annual mean error is shown for five and Climate Physics at The skill of a forecasting model can be lead times out to 120 hours. Errors have University College assessed by hindcasting, that is, computing clearly reduced over the period at all leads, London, UK, and lead what the model would have predicted in typically by 50%. The two main causes for scientist on the prior years had it been available at that this reduction are better real-time satellite TropicalStormRisk.com time. A hindcast for, say, the 1990 typhoon data, such as winds and atmospheric forecasting venture. Paul season is built using data up to and humidity, being assimilated into models, Rockett (bottom right) is including 1989. Figure 2 shows that and improvements in computational power. a Senior Research seasonal typhoon forecasts have useful skill For comparison, the skill of current Atlantic Fellow on the from early May. The TSR pre-season tropical cyclone track forecasts is typically TropicalStorm Risk.com forecast for 2000 performed particularly 10% better than the errors in Figure 3. venture. well. Issued on 26 May 2000, it called for 25 tropical storms, 14 typhoons, seven Future prospects intense typhoons and two typhoons The outlook is bright for further reaching Japan landfall. All of these innovations and improvement to both predictions proved correct. TSR’s further seasonal and track typhoon predictions. prediction of three tropical storms making Seasonal forecasting of annual typhoon landfall in Japan was the only one in error. activity is a new area of research. As such,

www.globalreinsurance.com