1 / 19

Precipitation Volatility and its Causes: Probability of Daily Extremes

Precipitation Volatility and its Causes: Probability of Daily Extremes. Alexander Gershunov 1 , Anna Panorska 2 and Tomasz Kozubowski 2 1 Climate, Atmospheric Science and Physical Oceanography (CASPO), Scripps Institution of Oceanography

paxton
Download Presentation

Precipitation Volatility and its Causes: Probability of Daily Extremes

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Precipitation Volatility and its Causes: Probability of Daily Extremes Alexander Gershunov1, Anna Panorska2 and Tomasz Kozubowski2 1Climate, Atmospheric Science and Physical Oceanography (CASPO), Scripps Institution of Oceanography 2Department of Mathematics and Statistics, University of Nevada Reno Panorska, A.K., A. Gershunov and T.J. Kozubowski, 2007: From diversity to volatility: Probability of daily precipitation extremes. In: A. Tsonis and J. Elsner (Eds.), Nonlinear Dynamics in Geosciences, Springer, New York, pp. 465-484. Kozubowski, T.J., A.K. Panorska, F. Qeadan and A. Gershunov, and D. Rominger, 2009: Testing exponentiality versus Pareto distribution via likelihood ratio. Communications in Statistics: Simulation and Computation, 38, 118-139.

  2. Motivations • A rigorous view of extreme precipitation events in climate requires a parametric approach. • Exponential tails may not be adequate to model probabilities of high-frequency precipitation extremes in volatile precipitation regions. We want to check how adequate or inadequate they are. • Indications are that heavy tailed distributions may be more appropriate for this task. • However, identifying and fitting heavy tails is difficult • Typical methods involve graphical analysis • There are many heavy-tailed distributions • Identification and parameter estimation needs to be automated

  3. Are precipitation extremes exponentially distributed? If not, what is a reasonable distribution? • Heavy-tailed modeling of extreme event probabilities • Heavy-tailed PDFs (power laws) allow for more extremes than traditional PDFs • Arise naturally as limit sums of random variables • Random variable X (e.g. precipitation) is "heavy tailed" if the probability (P) that it exceeds a value x is of power order x- for large x • P(X > x)  cx-, as x , where c,  > 0

  4. Approach • Peaks over threshold (POT) methodology • Three possible limiting PDFs for exceedances (approximations) • Balkema - de Haan - Pickands theorem (Balkema and de Haan 1974 and Pickands 1975) provides the limiting distribution of exceedances. The theorem says that when the threshold (u) increases, the distribution of the exceedance X[u] converges to a Generalized Pareto (GP) distribution. Any GP distribution has to be one of the following three kinds: exponential, Pareto or beta (finite, not applicable). So: no matter what the original distribution of X is, the exceedance X[u] over any threshold u is (approximately) one of only three distributions (effectively two). • Moreover, if original distribution has exponential tails, then the exceedance pdf will be exponential. If it has heavy tails, then the exceedance pdf will be Pareto. Using this result, we seek a statistic, a decision rule, to classify observed daily precipitation tails into exponential and heavy

  5. Exponential vs. Heavy Tails Ho: data comes from an exponential distribution, versus the alternative H1: data comes from a Pareto distribution. We approached this problem using ideas from the theory of likelihood ratio tests (Lehmann, 1997). The approach is to consider the ratio of the maxima of the likelihoods of the observed sample under the null (Pareto in the numerator) and alternative (exponential in the denominator) models. The logarithm of the likelihood ratio statistic is: Where is the observed sample (of excesses); and are the likelihood functions of the sample under Pareto and exponential hypotheses, respectively. We use a Pareto distribution with the survival function S(x) = P(X>x) = (1/(1+1/sα))α and exponential distribution with the survival function S(x) = P(X>x) = exp(-x/σ). In the Pareto case, the α parameter determines the thickness of its tail and is of primary importance. The scale parameter s is of secondary importance. In the exponential case σ is the scale parameter.

  6. Computation of L: the maximum likelihood procedure

  7. Statistical Properties of L Boxplots of simulated distributions of L. The first five boxplots were done using 10,000 observations of L from Pareto samples of size 1,000 with α varying from 0.5 (first boxplot) to 5 (second to the last boxplot). The last boxplot corresponds to 10,000 observations of L from exponential samples of size 1,000. The inset blows up the last two boxplots.

  8. Critical Values of L

  9. Log likelihood ratio: all data L values, absolute magnitude L values, Ho rejection certainty in % Log likelihood ratio (L) computed for daily excesses over local 75th percentile at each of the 560 stations. (a) Values close to zero, L <= 1 (blue and green x’s) represent approximately exponential tails, while yellow, red and black circles represent progressively heavier tails. (b) Level of confidence, (1 - γ)*100, for rejecting the null hypothesis (Ho) of exponential tails. Blue x’s represent exponential tails, green x’s represent stations at which the Ho cannot be rejected with reasonable (90%) confidence. Yellow and progressively redder circles represent stations at which Ho can be rejected with 90, 95, 98 and 99% confidence in favor of the Pareto alternative. For example, Ho can be rejected at 81% of stations with 95% confidence.

  10. Log likelihood ratio: seasonal data DJF MAM Four Seasons 40% heavy with 95% confidence 37% heavy with 95% confidence JJA SON 81% heavy 47% heavy with 95% confidence 51% heavy with 95% confidence

  11. Probability plots for selected stations Probability plots for excesses over threshold (75th percentile) at selected stations arranged in order of increasing L. Sorted observed excesses displayed in mm along the x-axis are plotted against the corresponding theoretical quantiles derived from the fitted exponential (blue x’s) and Pareto (red o’s) models.

  12. Precipitation Stats at Selected Stations

  13. Summary 1 • Diversity is the mother of volatility • Exponential tails are inadequate to model daily extremes in most regions of North America • Heavy tailed models are appropriate on theoretical and empirical grounds • These results can be directly extended to many climatic studies and applications • What about climate change? • How do climate models do?

  14. The Role of Topography Station Elevation % of precip falling in summer Tail structure

  15. The Role of Topography Aspect, Slope and tail structure Station Elevation % of precip falling in summer Tail structure

  16. The Role of Topography SLOPE and ASPECT distance from origin and direction from origin

  17. How do climate models do? Year-to-year correlation of modeled and observed max/median precipitation (a relative measure of volatility that is appropriate for model data) as a function of location (left) and as a function of slope and aspect (right). Contours show elevation in 250m (1000m) increments in black (blue). Statistics from 56 years of station observations (more than 400 stations) and a fine resolution (10km) downscaling of the NCEP/NCAR reanalysis (CaRD-10, performed by Kanamitsu and Kanamaru at SIO) over California were compared.

  18. Modeled Precipitation • Models don’t do heavy precipitation tails. • Untreated modeled daily extremes may be sensitive to climate only where observed precipitation tails are exponential. • Dynamical and/or statistical correction may be possible for other locations. • However, with climate change, non-stationarity is a big issue.

  19. Climate Change Challenges The choice between heavy tailed and exponentially tailed models is of a qualitative nature. The heavy tailed distributions have much larger high percentiles relative to the rest of the data values than the exponentially tailed ones. That implies that in places where heavy tailed models are appropriate, the future large events may be much larger than those observed up to date. The exponentially tailed models of precipitation will not be able to predict very large (relative to the observed data) events, because their mathematical properties do not allow such extremes. There is also the problem of non-stationarity…

More Related