Coherent functional forecasting 1 Coherent functional forecasts of mortality rates and life expectancy Rob J Hyndman Business & Economic Forecasting Unit Joint work with Farah Yasmeen (Monash) and Heather Booth (ANU) Coherent functional forecasting 2 Mortality rates −4 −6 −8 Log death rate −2 Australia male mortality: 1950 0 20 40 60 Age 80 100 Coherent functional forecasting 3 Mortality rates −4 −6 −8 Log death rate −2 Australia female mortality: 1950 0 20 40 60 Age 80 100 Coherent functional forecasting The problem Let ft,j (x) be the smoothed mortality rate for age x in group j in year t. 4 Coherent functional forecasting The problem Let ft,j (x) be the smoothed mortality rate for age x in group j in year t. Groups may be males and females. 4 Coherent functional forecasting The problem Let ft,j (x) be the smoothed mortality rate for age x in group j in year t. Groups may be males and females. Groups may be states within a country. 4 Coherent functional forecasting The problem Let ft,j (x) be the smoothed mortality rate for age x in group j in year t. Groups may be males and females. Groups may be states within a country. Expected that groups will behave similarly. 4 Coherent functional forecasting The problem Let ft,j (x) be the smoothed mortality rate for age x in group j in year t. Groups may be males and females. Groups may be states within a country. Expected that groups will behave similarly. We want to forecast whole curve ft,j (x) for future years. 4 Coherent functional forecasting The problem Let ft,j (x) be the smoothed mortality rate for age x in group j in year t. Groups may be males and females. Groups may be states within a country. Expected that groups will behave similarly. We want to forecast whole curve ft,j (x) for future years. Coherent forecasts do not diverge over time. 4 Coherent functional forecasting The problem Let ft,j (x) be the smoothed mortality rate for age x in group j in year t. Groups may be males and females. Groups may be states within a country. Expected that groups will behave similarly. We want to forecast whole curve ft,j (x) for future years. Coherent forecasts do not diverge over time. Existing functional models do not impose coherence. 4 Coherent functional forecasting 5 Functional time series model (Hyndman and Ullah, CSDA, 2007) log[ft,j (x)] = µj (x) + K X k =1 iid where et,j (x) ∼ N(0, v(x)). βt,j ,k φk ,j (x) + et,j (x) Coherent functional forecasting 5 Functional time series model (Hyndman and Ullah, CSDA, 2007) log[ft,j (x)] = µj (x) + K X βt,j ,k φk ,j (x) + et,j (x) k =1 iid where et,j (x) ∼ N(0, v(x)). 1 Estimate smooth functions ft,j (x) using nonparametric regression. Coherent functional forecasting 5 Functional time series model (Hyndman and Ullah, CSDA, 2007) log[ft,j (x)] = µj (x) + K X βt,j ,k φk ,j (x) + et,j (x) k =1 iid where et,j (x) ∼ N(0, v(x)). 1 2 Estimate smooth functions ft,j (x) using nonparametric regression. Estimate µj (x) as mean log[ft,j (x)] across years. Coherent functional forecasting 5 Functional time series model (Hyndman and Ullah, CSDA, 2007) log[ft,j (x)] = µj (x) + K X βt,j ,k φk ,j (x) + et,j (x) k =1 iid where et,j (x) ∼ N(0, v(x)). 1 2 3 Estimate smooth functions ft,j (x) using nonparametric regression. Estimate µj (x) as mean log[ft,j (x)] across years. Estimate βt,j ,k and φk ,j (x) using functional principal components. Coherent functional forecasting 5 Functional time series model (Hyndman and Ullah, CSDA, 2007) log[ft,j (x)] = µj (x) + K X βt,j ,k φk ,j (x) + et,j (x) k =1 iid where et,j (x) ∼ N(0, v(x)). 1 2 3 4 Estimate smooth functions ft,j (x) using nonparametric regression. Estimate µj (x) as mean log[ft,j (x)] across years. Estimate βt,j ,k and φk ,j (x) using functional principal components. Forecast βt,j ,k using time series models. Coherent functional forecasting 5 Functional time series model (Hyndman and Ullah, CSDA, 2007) log[ft,j (x)] = µj (x) + K X βt,j ,k φk ,j (x) + et,j (x) k =1 iid where et,j (x) ∼ N(0, v(x)). 1 2 3 4 5 Estimate smooth functions ft,j (x) using nonparametric regression. Estimate µj (x) as mean log[ft,j (x)] across years. Estimate βt,j ,k and φk ,j (x) using functional principal components. Forecast βt,j ,k using time series models. Put it all together to get forecasts of ft,j (x). Coherent functional forecasting 6 0.4 0.2 φ3(x) −0.2 0 20 40 60 80 0 Age 20 40 60 80 0 Age 20 40 60 80 Age 1970 1990 Year 1950 0.2 0.0 βt3 −0.2 −0.4 βt2 1950 −0.6 −0.4 −0.2 −2 −1 βt1 0 0.0 1 0.2 2 Age −0.4 −0.6 0.1 20 40 60 80 −3 0 0.0 φ2(x) −0.2 0.0 φ1(x) 0.2 −4 −6 −8 µM(x) 0.3 0.2 −2 0.4 0.4 Male fts model 1970 1990 Year 1950 1970 1990 Year Coherent functional forecasting 7 φ3(x) −0.4 0 20 40 60 80 0 Age βt2 −0.1 0.0 0.1 2 1 0 1950 1970 1990 Year 1950 1970 1990 Year −0.4 −0.3 −2 −1 20 40 60 80 0.0 0.1 0.2 Age 0.2 Age βt1 −0.2 0.0 20 40 60 80 βt3 0 Age −0.2 20 40 60 80 −3 0 −0.4 0.1 −0.2 0.0 φ2(x) φ1(x) 0.2 −4 −6 −8 µF(x) 0.3 0.2 −2 0.2 0.4 0.4 Female fts model 1950 1970 1990 Year Coherent functional forecasting 8 Forecasting the coefficients log[ft,j (x)] = µj (x) + K X βt,j ,k φk ,j (x) + et,j (x) k =1 We use ARIMA models for each coefficient {β1,j ,k , . . . , βn,j ,k }. Coherent functional forecasting 8 Forecasting the coefficients log[ft,j (x)] = µj (x) + K X βt,j ,k φk ,j (x) + et,j (x) k =1 We use ARIMA models for each coefficient {β1,j ,k , . . . , βn,j ,k }. The ARIMA models are non-stationary for the first few coefficients (k = 1, 2) Coherent functional forecasting 8 Forecasting the coefficients log[ft,j (x)] = µj (x) + K X βt,j ,k φk ,j (x) + et,j (x) k =1 We use ARIMA models for each coefficient {β1,j ,k , . . . , βn,j ,k }. The ARIMA models are non-stationary for the first few coefficients (k = 1, 2) Non-stationary ARIMA forecasts will diverge. Hence the mortality forecasts are not coherent. Coherent functional forecasting 9 0.4 0.2 φ3(x) −0.2 20 40 60 80 0 20 40 60 80 0.2 1 βt3 −0.2 βt2 −0.4 −1 −2 2000 Year 20 40 60 80 Age 0 −5 βt1 −10 1960 0 Age 0 Age 2040 0.0 0 Age −0.4 −0.6 0.1 20 40 60 80 −15 0 0.0 φ2(x) −0.2 0.0 φ1(x) 0.2 −4 −6 −8 µM(x) 0.3 0.2 −2 0.4 0.4 Male fts model 1960 2000 Year 2040 1960 2000 Year 2040 Coherent functional forecasting 10 20 40 60 80 φ3(x) −0.4 0 20 40 60 80 Age 0.0 βt3 −0.2 βt2 −0.2 −0.4 −0.6 2000 Year 20 40 60 80 0.0 0.1 0.2 2 0 −2 −4 −6 βt1 0 Age 0.2 Age 1960 −0.2 0.0 0 Age 2040 1960 2000 Year 2040 −0.4 20 40 60 80 −8 0 −0.4 0.1 −0.2 0.0 φ2(x) φ1(x) 0.2 −4 −6 −8 µF(x) 0.3 0.2 −2 0.2 0.4 0.4 Female fts model 1960 2000 Year 2040 Coherent functional forecasting 11 Australian mortality forecasts −6 −4 0 20 40 60 Age 80 100 −12 −10 −8 Log death rate −6 −8 −10 −12 Log death rate −4 −2 (b) Females −2 (a) Males 0 20 40 60 Age 80 100 Coherent functional forecasting Mortality product and ratios Key idea Model the geometric mean and the mortality ratio instead of the individual rates for each sex separately. 12 Coherent functional forecasting Mortality product and ratios Key idea Model the geometric mean and the mortality ratio instead of the individual rates for each sex separately. q q . pt (x) = ft,M (x)ft,F (x) and rt (x) = ft,M (x) ft,F (x). 12 Coherent functional forecasting Mortality product and ratios Key idea Model the geometric mean and the mortality ratio instead of the individual rates for each sex separately. q q . pt (x) = ft,M (x)ft,F (x) and rt (x) = ft,M (x) ft,F (x). Product and ratio are approximately independent 12 Coherent functional forecasting Mortality product and ratios Key idea Model the geometric mean and the mortality ratio instead of the individual rates for each sex separately. q q . pt (x) = ft,M (x)ft,F (x) and rt (x) = ft,M (x) ft,F (x). Product and ratio are approximately independent Ratio should be stationary (for coherence) but product can be non-stationary. 12 Coherent functional forecasting 13 Mortality rates −4 −6 −8 Log death rate −2 Australia gmean mortality: 1950 0 20 40 60 Age 80 100 Coherent functional forecasting 14 Mortality rates 3.0 2.5 2.0 1.5 1.0 sex ratio: M/F 3.5 4.0 Australia mortality sex ratio 1950 0 20 40 60 Age 80 100 Coherent functional forecasting 15 Model product and ratios q q . pt (x) = ft,M (x)ft,F (x) and rt (x) = ft,M (x) ft,F (x). log[pt (x)] = µp (x) + log[rt (x)] = µr (x) + K X k =1 L X `=1 βt,k φk (x) + et (x) γt,` ψ` (x) + wt (x). Coherent functional forecasting 15 Model product and ratios q q . pt (x) = ft,M (x)ft,F (x) and rt (x) = ft,M (x) ft,F (x). log[pt (x)] = µp (x) + log[rt (x)] = µr (x) + K X k =1 L X βt,k φk (x) + et (x) γt,` ψ` (x) + wt (x). `=1 {γt,` } restricted to be stationary processes: either ARMA(p, q) or ARFIMA(p, d , q). Coherent functional forecasting 15 Model product and ratios q q . pt (x) = ft,M (x)ft,F (x) and rt (x) = ft,M (x) ft,F (x). log[pt (x)] = µp (x) + log[rt (x)] = µr (x) + K X k =1 L X βt,k φk (x) + et (x) γt,` ψ` (x) + wt (x). `=1 {γt,` } restricted to be stationary processes: either ARMA(p, q) or ARFIMA(p, d , q). No restrictions for βt,1 , . . . , βt,K . Coherent functional forecasting 15 Model product and ratios q q . pt (x) = ft,M (x)ft,F (x) and rt (x) = ft,M (x) ft,F (x). log[pt (x)] = µp (x) + log[rt (x)] = µr (x) + K X k =1 L X βt,k φk (x) + et (x) γt,` ψ` (x) + wt (x). `=1 {γt,` } restricted to be stationary processes: either ARMA(p, q) or ARFIMA(p, d , q). No restrictions for βt,1 , . . . , βt,K . Forecasts: fn+h |n,M (x) = pn+h |n (x)r.n+h |n (x) fn+h |n,F (x) = pn+h |n (x) rn+h |n (x). Coherent functional forecasting 16 0 20 40 60 80 0.2 20 40 60 80 0 20 40 60 80 Age Age 1 0.2 2 βt3 0 βt2 −2 −0.2 −1 0.0 0.1 0 −2 −4 −6 βt1 0.0 φ3(x) −0.4 0 Age 0.3 Age −0.2 0.0 φ2(x) 0.1 20 40 60 80 −8 0 −0.4 −0.2 φ1(x) 0.2 −4 −6 −8 µP(x) 0.3 0.2 −2 0.4 0.4 Product model 1960 2000 Year 2040 1960 2000 Year 2040 1960 2000 Year 2040 Coherent functional forecasting 17 0.6 φ3(x) 0.0 0.0 φ2(x) 0.2 0.2 φ1(x) 0.4 0.4 0.4 0.5 0.4 0.3 −0.2 −0.4 0.0 0.2 20 40 60 80 0 20 40 60 80 0 20 40 60 80 Age 0 20 40 60 80 Age Age 0.10 βt3 0.0 βt2 1960 2000 Year 2040 −0.10 −0.2 −0.2 −0.1 βt1 0.0 0.1 0.2 0.2 Age −0.4 0 0.00 0.1 µR(x) 0.8 Ratio model 1960 2000 Year 2040 1960 2000 Year 2040 Coherent functional forecasting 18 −4 −6 −8 −10 Log of geometric mean death rate −2 Product forecasts 0 20 40 60 Age 80 100 Coherent functional forecasting 19 3.0 2.5 2.0 1.5 1.0 Sex ratio: M/F 3.5 4.0 Ratio forecasts 0 20 40 60 Age 80 100 Coherent functional forecasting 20 Coherent forecasts −4 −10 −8 −6 Log death rate −6 −8 −10 Log death rate −4 −2 (b) Females −2 (a) Males 0 20 40 60 Age 80 100 0 20 40 60 Age 80 100 Coherent functional forecasting 21 Ratio forecasts 4 3 1 2 Sex ratio of rates: M/F 3 2 1 Sex ratio of rates: M/F 4 5 Coherent forecasts 5 Independent forecasts 0 20 40 60 Age 80 100 0 20 40 60 Age 80 100 Coherent functional forecasting 22 Life expectancy difference: F−M 8 Life expectancy forecasts 4 Number of years 80 2 75 70 Age 85 6 90 95 Life expectancy forecasts 1960 2000 Year 2040 1960 2000 Year 2040 Coherent functional forecasting 23 Coherent forecasts for J groups and pt (x) = [ft,1 (x)ft,2 (x) · · · ft,J (x)]1/J . rt,j (x) = ft,j (x) pt (x), Coherent functional forecasting 23 Coherent forecasts for J groups and pt (x) = [ft,1 (x)ft,2 (x) · · · ft,J (x)]1/J . rt,j (x) = ft,j (x) pt (x), log[pt (x)] = µp (x) + K X βt,k φk (x) + et (x) k =1 L X log[rt,j (x)] = µr,j (x) + l =1 γt,l ,j ψl ,j (x) + wt,j (x). Coherent functional forecasting 23 Coherent forecasts for J groups and pt (x) = [ft,1 (x)ft,2 (x) · · · ft,J (x)]1/J . rt,j (x) = ft,j (x) pt (x), log[pt (x)] = µp (x) + K X βt,k φk (x) + et (x) k =1 L X log[rt,j (x)] = µr,j (x) + l =1 pt (x) and all rt,j (x) are approximately independent. γt,l ,j ψl ,j (x) + wt,j (x). Coherent functional forecasting 23 Coherent forecasts for J groups and pt (x) = [ft,1 (x)ft,2 (x) · · · ft,J (x)]1/J . rt,j (x) = ft,j (x) pt (x), log[pt (x)] = µp (x) + K X βt,k φk (x) + et (x) k =1 L X log[rt,j (x)] = µr,j (x) + γt,l ,j ψl ,j (x) + wt,j (x). l =1 pt (x) and all rt,j (x) are approximately independent. Ratios satisfy constraint rt,1 (x)rt,2 (x) · · · rt,J (x) = 1. Coherent functional forecasting 23 Coherent forecasts for J groups and pt (x) = [ft,1 (x)ft,2 (x) · · · ft,J (x)]1/J . rt,j (x) = ft,j (x) pt (x), log[pt (x)] = µp (x) + K X βt,k φk (x) + et (x) k =1 L X log[rt,j (x)] = µr,j (x) + γt,l ,j ψl ,j (x) + wt,j (x). l =1 pt (x) and all rt,j (x) are approximately independent. Ratios satisfy constraint rt,1 (x)rt,2 (x) · · · rt,J (x) = 1. log[ft,j (x)] = log[pt (x)rt,j (x)] Coherent functional forecasting 24 Coherent forecasts for J groups log[ft,j (x)] = log[pt (x)rt,j (x)] = log[pt (x)] + log[rt,j ] K L X X = µj (x) + βt,k φk (x) + γt,`,j ψ`,j (x) + zt,j (x) k =1 `=1 Coherent functional forecasting 24 Coherent forecasts for J groups log[ft,j (x)] = log[pt (x)rt,j (x)] = log[pt (x)] + log[rt,j ] K L X X = µj (x) + βt,k φk (x) + γt,`,j ψ`,j (x) + zt,j (x) k =1 `=1 µj (x) = µp (x) + µr,j (x) is group mean Coherent functional forecasting 24 Coherent forecasts for J groups log[ft,j (x)] = log[pt (x)rt,j (x)] = log[pt (x)] + log[rt,j ] K L X X = µj (x) + βt,k φk (x) + γt,`,j ψ`,j (x) + zt,j (x) k =1 `=1 µj (x) = µp (x) + µr,j (x) is group mean zt,j (x) = et (x) + wt,j (x) is error term. Coherent functional forecasting 24 Coherent forecasts for J groups log[ft,j (x)] = log[pt (x)rt,j (x)] = log[pt (x)] + log[rt,j ] K L X X = µj (x) + βt,k φk (x) + γt,`,j ψ`,j (x) + zt,j (x) k =1 `=1 µj (x) = µp (x) + µr,j (x) is group mean zt,j (x) = et (x) + wt,j (x) is error term. {γt,` } restricted to be stationary processes: either ARMA(p, q) or ARFIMA(p, d , q). Coherent functional forecasting 24 Coherent forecasts for J groups log[ft,j (x)] = log[pt (x)rt,j (x)] = log[pt (x)] + log[rt,j ] K L X X = µj (x) + βt,k φk (x) + γt,`,j ψ`,j (x) + zt,j (x) k =1 `=1 µj (x) = µp (x) + µr,j (x) is group mean zt,j (x) = et (x) + wt,j (x) is error term. {γt,` } restricted to be stationary processes: either ARMA(p, q) or ARFIMA(p, d , q). No restrictions for βt,1 , . . . , βt,K . Coherent functional forecasting Li-Lee method Li & Lee (Demography, 2005) method is a special case of our approach. ft,j (x) = µj (x) + βt φ(x) + γt,j ψj (x) + et,j (x) where f is unsmoothed log mortality rate, βt is a random walk with drift and γt,j is AR(1) process. No smoothing. 25 Coherent functional forecasting Li-Lee method Li & Lee (Demography, 2005) method is a special case of our approach. ft,j (x) = µj (x) + βt φ(x) + γt,j ψj (x) + et,j (x) where f is unsmoothed log mortality rate, βt is a random walk with drift and γt,j is AR(1) process. No smoothing. Only one basis function for each part, 25 Coherent functional forecasting Li-Lee method Li & Lee (Demography, 2005) method is a special case of our approach. ft,j (x) = µj (x) + βt φ(x) + γt,j ψj (x) + et,j (x) where f is unsmoothed log mortality rate, βt is a random walk with drift and γt,j is AR(1) process. No smoothing. Only one basis function for each part, Random walk with drift very limiting. 25 Coherent functional forecasting Li-Lee method Li & Lee (Demography, 2005) method is a special case of our approach. ft,j (x) = µj (x) + βt φ(x) + γt,j ψj (x) + et,j (x) where f is unsmoothed log mortality rate, βt is a random walk with drift and γt,j is AR(1) process. No smoothing. Only one basis function for each part, Random walk with drift very limiting. AR(1) very limiting. 25 Coherent functional forecasting Li-Lee method Li & Lee (Demography, 2005) method is a special case of our approach. ft,j (x) = µj (x) + βt φ(x) + γt,j ψj (x) + et,j (x) where f is unsmoothed log mortality rate, βt is a random walk with drift and γt,j is AR(1) process. No smoothing. Only one basis function for each part, Random walk with drift very limiting. AR(1) very limiting. The γt,j coefficients will be highly correlated with each other, and so independent models are not appropriate. 25 Coherent functional forecasting Some conclusions New, automatic, flexible method for coherent forecasting of groups of functional time series. 26 Coherent functional forecasting Some conclusions New, automatic, flexible method for coherent forecasting of groups of functional time series. Based on geometric means and ratios, so interpretable results. 26 Coherent functional forecasting Some conclusions New, automatic, flexible method for coherent forecasting of groups of functional time series. Based on geometric means and ratios, so interpretable results. More general and flexible than existing methods. 26 Coherent functional forecasting Some conclusions New, automatic, flexible method for coherent forecasting of groups of functional time series. Based on geometric means and ratios, so interpretable results. More general and flexible than existing methods. Easy to compute prediction intervals. 26