Piet de Jong
Transcription
Piet de Jong
Inferring and predicting global temperature trends Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney June 29, 2012 Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Temperature records – GISS, CRU and UAH 15 C 14 13 12 1840 1860 1880 1900 1920 1940 1960 1980 2000 2020 year 1562 + 1922 + 375 = 3859 monthly observations Many other time series like this available of differing lengths, frequency, relatedness (tree thickness), missing data properties, validity (urban heat islands), comprehensiveness etc. (Selection bias) Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Temperature trends Monthly – with northern/southern hemisphere seasonality issues removed Different starts for the 3 series – all end at 2010 All “standardized” on same (arbitrary) level – even though they are centred differently (see GLS later) Noisy – but each appears to suggest (global) warming Decreasing volatility? Evidence of this sort is basis for $trillions proposed taxes/expenditure on climate change mitigation strategies What really is the trend/slope out into the future and associated confidence intervals? Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Overview Time series analysis done on such “climatic” time series is typically inadequate an open to much criticism Many, many people “laying in” on such analyses – similar to stock market “chartism” See for example the various websites: climateaudit.org, realclimate.org etc etc for the bitterness of the controversies. Most “statistical” analyses (even by the “climate scientists”) ignore the time series literature in particular as well as the broader statistical literature (significance, selection biases, preprocessing issues often ignored/swept away). We aim to partly critique existing approaches and provide a more “proper” approach. “Proper” approach is based on the ideas and methods initiated, motivated, inspired and developed by Andrew Harvey. Area is controversial, politicised, at times bitter. Warrants attention from proper time series analysts. Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends “Mannian” smoothing – after Mann (2004) This may be typical what is goes in “climatic time series analysis.” To smooth a climate series: Use a “Butterworth” filter to smooth the time series Where the filter runs into either end of the series just “mirror” the time series “horizontally” about t = 0 or t = n. To “preserve trend/slope,” mirror the mirrored series “vertically” about y = yn . (“Double mirroring”) Result is a smoothing procedure, unhampered by end effects, that “properly resolves” the trend Why Butterworth? What, are the statistical properties of such a procedure? Voodoo smoothing? ỹn = yn always! How do these smoothing procedures relate to “proper” time series procedures? Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Example series: Northern Hemisphere Temperature Index – Mann(2008) Annual: 1850 – Current 1 correlation 0.5 C 0 -0.5 1850 1900 1950 0.5 0 -0.5 2000 0 5 year 15 20 10 15 20 1 correlation 1 correlation 10 lag 0.5 0 -0.5 0 5 10 15 20 lag Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney 0.5 0 -0.5 0 5 lag Inferring and predicting global temperature trends Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Harvey & Trimbur (2003) Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Connection between Butterworth and Components Model – summary Components Model m – order of differencing q = σζ2 /σ�2 Butterworth m ”Slope” of the gain function Determine window half–width: 40 years? λ0 equal to reciprocal of window half width Determine q = {2 sin(λ0 /2)}2m Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Note: End effects handled properly above but “improperly” by Mann etc Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Busetti Harvey (2007) Model Suppose stochastic trend (m = 2) + AR(1)+WN yt = µ t + a t + λ � t , st+1 = st + ηt , µt+1 = µt + st , at+1 = ρ at + θ νt , �t ∼ (0, σ 2 ) ηt , νt ∼ (0, σ 2 ) Note: λ and θ are “noise–to–signal” parameters. σ 2 is scale parameter: not of interest Imagine model holds for 3 series with µt , at “common” and �t r r “specific” with λ equal to θ times e 1 , e a + b t or e 3 µt to be inferred/predicted: inference/prediction is “pooled” Each series “shifted” with “starting conditions” to “level” each series on the common µt . How? Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Forecasts + Error Bounds based on KF/SF slope level 16 C C/month 17 0.005 0 15 14 13 1850 1900 1950 2000 2050 year 1850 1900 1950 2000 2050 year Pooled estimator: > 3 series unlikely to narrow bounds No distinction between “inferring” and “prediction” End effects handled “properly” through use of KF/SF MIT Professor Richard Lindzen: “no statistically significant warming since 1995” MIT President Susan Hockfield: climate change “accelerating.” Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Technicalities: starts and shifts KF/SF dealt with all estimation/prediction issues including the problem of “end effects” (vs Mannian smoothing) KF/SF dealt with different starting points through use of different starting conditions: Estimates of shifts and initial conditions estimate std. dev Shift GISS UAH 14.0753 14.3381 0.0035 0.0070 CRU Jan. 1850 level Jan. 1850 slope 13.8225 -0.0004 0.0630 0.0006 Estimation in terms of ln(θ) and ln(λi ) = ln(θ) + ri , i = 1, 3 and r2 = a + bt allowing for changing volatility. Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Model parameter estimates estimate std dev. correlation matrix AR ln(θ) 7.96 0.21 1.00 -0.08 -0.09 0.04 -0.06 0.06 GISS r1 -0.86 0.06 -0.08 1.00 0.54 -0.32 0.16 0.43 parameter vector ψ CRU UAH a b r3 0.83 -0.87 -0.22 0.04 0.03 0.03 -0.09 0.04 -0.06 0.54 -0.32 0.16 1.00 -0.75 0.28 -0.75 1.00 -0.11 0.28 -0.11 1.00 0.22 -0.10 0.06 AR coeff. ρ 0.63 0.02 0.06 0.43 0.22 -0.10 0.06 1.00 Estimates and cov matrix derived from KF/SF based likelihood evaluation Parameter estimate uncertainty used to simulate future scenarios Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends 2060 temperature Predicted 2060 global temperature 18 18 17 17 16 16 15 15 14 14 13 13 1.02 1.04 1.06 1.08 1.1 1.12 600 c : change in temp 0 0.2 0.4 0.6 0.8 1 exceedence probability Figure: Forecast February 2060 temperature. The left panel plots the forecast versus the estimate of the 2010 slope. Different points correspond to different choices of ψ. The right panel indicates exceedence probabilities with the flatter curve assuming ψ = ψ̂ and the steeper curve factoring in the uncertainty regarding ψ. Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Is the model adequate? – Stress tests slope 1 1 0.8 0.8 p-value p-value level 0.6 0.4 0.2 0.6 0.4 0.2 0 0 1850 1900 1950 2000 year 1850 1900 1950 2000 year Figure: Left panel display p–values of estimated shocks introduced to the level while the right hand side is p–values for the slope. Largest p–values for shocks to st all occur in the period 1963–1977 indicating the model is under most stress from rapidly increasing temperatures during this period. Little evidence that shocks are warranted to the slope post 1977. Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends Conclusions – “real” time series analysts should step up to the plate Professor Andrew Harvey has been instrumental in inspiring, developing, applying, exploring, justifying a very useful and powerful approach to time series analysis rooted in the statistical paradigm This approach is largely if not completely ignored by the “climate science community” The approach is completely applicable and useful for the extensive set of climate time series More should be done to “muscle in” on this territory. Craig Ansley, NZ & Piet de Jong , Macquarie University, Sydney Inferring and predicting global temperature trends