Forecasts

SST-field forecast

HiDyn-Model forecasts of SST anomalies for the Tropical Pacific Region are available from August 1980 to the present. Forecasts are based on data seven months prior to the forecast back to January-1970 data. For example, the December-1985 forecast uses May-1985 data back to January-1970 data.

SST-field forecast animations

HiDyn-Model forecasts of SST anomalies for the Tropical Pacific Region are animated. In these animations, each frame is based on data 7 months prior to the forecast back to January-1970 data. For example, the December-1985 forecast uses May-1985 data back to January-1970 data; the next frame of the animation, namely for January 1986, is based on all the data used to forecast December 1985, plus the extra month of data for June 1985.

Warm/Normal/Cool Regimes

The HiDyn Model developed in "Long-Lead Prediction of Pacific SSTs via Bayesian Dynamic Modeling" by L.M. Berliner, C.K. Wikle, and N. Cressie (2000), Journal of Climate, 13, 3953-3968, uses current values of SST anomalies, the Southern Oscillation Index (SOI), and a summary of westerly surface-wind bursts as predictor variables. Rather than viewing the prediction as restricted to a single model, several statistical prediction models are developed. These models condition on the current regime (Warm, Normal, or Cool), classified according to the current value of SOI, and then provide probabilistic forecasts of the future regimes (Warm, Normal, or Cool) seven months later. The probabilities of the future regimes are estimated based on the current SOI and wind-burst statistics. This model was trained on monthly data beginning in 1970.

The HiDyn-Model output is the predictive distribution for SST anomalies in the Tropical Pacific Region, with a seven-month lead. Key summaries of this distribution include (i) probabilities of each of the three temperature-regime states; and (ii) SST mean-field estimates for each temperature regime. This page shows (i) and (ii). When a probability-weighted average of the Warm, Normal, and Cool mean-field estimates in (ii) is taken (with probabilities given by (i)), we obtain (iii) a combined mean-field estimate that yields the SST-field forecasts seen on most of the other ENSO webpages.

Warm/Normal/Cool Forecast

HiDyn-Model forecasts of SST anomalies for the Tropical Pacific Region are available from August 1980 to the present. Forecasts are based on data seven months prior to the forecast back to January-1970 data. For example, the December-1985 forecast uses May-1985 data back to January-1970 data.

Warm/Normal/Cool Regime - Animation

Not Available

Regime HiDyn-Model forecasts of SST anomalies for August 2012 to July 2013 (Frame delay = 1.00 sec). Probabilities of the regime forecast being in Warm, Normal, or Cool regimes are plotted in barycentric co-ordinates.

Warm/Normal/Cool Regimes

The HiDyn Model developed in "Long-Lead Prediction of Pacific SSTs via Bayesian Dynamic Modeling" by L.M. Berliner, C.K. Wikle, and N. Cressie (2000), Journal of Climate, 13, 3953-3968, uses current values of SST anomalies, the Southern Oscillation Index (SOI), and a summary of westerly surface-wind bursts as predictor variables. Rather than viewing the prediction as restricted to a single model, several linked statistical prediction models are developed. These models condition on the current regime (Warm, Normal, or Cool), classified according to the current value of SOI, and then provide probabilistic forecasts of future regimes (Warm, Normal, or Cool) seven months later. The probabilities of the future regimes are estimated based on the current SOI and wind-burst statistics. This model was trained on monthly data begining in 1970.

The HiDyn-Model output is the predictive distribution for SST anomalies in the Tropical Pacific, with a seven-month lead. Key summaries of this distribution include (i) probabilities of each of the three temperature-regime states; and (ii) SST mean-field estimates for each temperature regime. This page shows (i) and (ii). When a probability-weighted average of the Warm, Normal, and Cool mean-field estimates in (ii) is taken (with probabilities given by (i)), we obtain (iii) a combined mean-field estimate that yields the SST-field forecasts seen on most of the other ENSO webpages.

Warm/Normal/Cool Forecast Animations

In these animations, each frame is based on data 7 months prior to the forecast back to January-1970 data. For example, the December-1985 forecast uses May-1985 data back to January-1970 data; the next frame of the animation, namely for January 1986, is based on all the data used to forecast December 1985, plus the extra month of data for June 1985.

Nino 3.4 Forecast

The Nino 3.4 Region is illustrated on the ENSO Home Page. HiDyn-Model forecasts of Nino 3.4 anomalies are available from August 1980 to the present. Forecasts are based on data seven months prior to the forecast back to January-1970 data. For example, the December-1985 forecast uses May-1985 data back to January-1970 data.

Nino 3.4 Forecast

The Nino 3.4 Region is illustrated on the ENSO Home Page. Nino 3.4 forecasts are available from August 1980 to the present. Forecasts are based on data seven months prior to the forecast back to January-1970 data. For example, the December-1985 forecast uses May-1985 data back to January-1970 data.

Observed SST Anomaly Map

HiDyn-Model Forecast Map of SST Anomalies 

HiDyn-Model forecast of SST anomalies for December 2012 based on May-2012 data back to January-1970 data. The Nino 3.4 region is outlined.

SST Forecast Comparisons

Observed SST anomalies are available for forecast comparisons up to the most recent month of data. HiDyn-Model forecasts for Nino 3.4 anomalies are available from August 1980 up to seven months beyond the most recent month of data. Comparisons are only possible for a selected month for which both the observed ("true") and the forecast SST anomalies are available. A measure of performance of the forecast map is, Performance = ave{(forecast - observed)²}, where the average is taken over all pixels in the Nino 3.4 Region.

Two forecasts, Forecast A and Forecast B, can be compared by the Relative Performance (RP) measure,

Note that RP(A:B) > 0 indicates that Forecast A performs better than Forecast B.

The HiDyn-Model forecast is compared with two easy-to-compute forecast procedures, namely the Persistence forecast and the Climatology forecast. The Persistence forecast, for a selected month, consists of the observed SST anomalies in the Nino 3.4 Region 7 months earlier. For example, the Persistence forecast for October 1998 consists of the observed anomalies for March 1998. The Climatology forecast is obtained from long-term (typically 30 years) averages of data for each month of the year. Because the data are anomalies, this forecast for any month and any pixel in the Nino 3.4 Region (in fact for any pixel in the Tropical Pacific Region) is 0. Note that to make comparisons of the HiDyn Model's forecasts to another forecast procedure using the RP measure, one must have pixel-level forecasts from the other procedure available in the Nino 3.4 Region.

The performance (or skill) of forecasts can also be compared through their overall correlation with the observed values,

where

and A^* = ave(Forecast A) and O^* = ave(Observed).

Note that C(A)>C(B) indicates that Forecast A has better performance than Forecast B.

The following table shows C(HiDyn), the correlation coefficient of anomalies of the 7-month-lead HiDyn-Model forecast and C(Persistence), the 7-month-lead Persistence forecast, where the correlations are calculated separately for two 15-year-long periods, 1987-2001 and 1992-2006.

Correlation coefficients for other ENSO forecast models in the Nino 3.4 Region can be found in van Oldenborgh et al. (2005). The correlation coefficients listed in the table above correspond to a "+6" monthly Nino-3.4 forecast in the terminology of van Oldenborgh et al. (2005). However, they do not go beyond "+5" monthly Nino 3.4 forecasts, so our results and theirs cannot be directly compared.

Performance measures based on correlation coefficients have the disadvantage that a forecast could be biased but still have a maximum skill of 1 (a correlation coefficient does not distinguish between forecast=observed and forecast=0.5*observed +1°C). The HiDyn Model is built to forecast well at the 1° × 1° pixel scale in the whole Tropical Pacific Ocean, not just at a regional scale on Nino 3.4. Another skill measure that might be tried is whether the warm/normal/cold regimes of the forecast are correct (These regimes are defined in Berliner et al., 2000). The HiDyn Model should have very high performance in this regard, because of the way it is built.

Reference:

L.M. Berliner, C.K. Wikle, and N. Cressie, 2000. Long-Lead Prediction of Pacific SSTs via Bayesian Dynamic Modeling. Journal of Climate, 13, 3953-3968.

G.J. van Oldenborgh, M.A. Balmaseda, L. Ferranti, T.N. Stockdale, and D.L.T. Anderson, 2005. Did the ECMWF Seasonal Forecast Model Outperform Statistical ENSO Forecast Models over the Last 15 Years? Journal of Climate, 18, 3240-3249.

Relative Performance (RP) of HiDyn-Model forecast relative to Persistence forecast (red plus) and to Climatology forecast (blue cross), for 2012 to 2012. RP > 0 indicates HiDyn Model performs better.

SST Forecast Comparisons

Observed SST anomalies are available for forecast comparisons up to the most recent month of data. HiDyn-Model forecasts for Nino 3.4 anomalies are available from August 1980 up to seven months beyond the most recent month of data. Comparisons are only possible for a selected month for which both the observed ("true") and the forecast SST anomalies are available. A measure of performance of the forecast map is, Performance = ave{(forecast - observed)²}, where the average is taken over all pixels in the Nino 3.4 Region.

Two forecasts, Forecast A and Forecast B, can be compared by the Relative Performance (RP) measure,

Note that RP(A:B) > 0 indicates that Forecast A performs better than Forecast B.

The HiDyn-Model forecast is compared with two easy-to-compute forecast procedures, namely the Persistence forecast and the Climatology forecast. The Persistence forecast, for a selected month, consists of the observed SST anomalies in the Nino 3.4 Region 7 months earlier. For example, the Persistence forecast for October 1998 consists of the observed anomalies for March 1998. The Climatology forecast is obtained from long-term (typically 30 years) averages of data for each month of the year. Because the data are anomalies, this forecast for any month and any pixel in the Nino 3.4 Region (in fact for any pixel in the Tropical Pacific Region) is 0. Note that to make comparisons of the HiDyn Model's forecasts to another forecast procedure using the RP measure, one must have pixel-level forecasts from the other procedure available in the Nino 3.4 Region.

The performance (or skill) of forecasts can also be compared through their overall correlation with the observed values,

where

and A^* = ave(Forecast A) and O^* = ave(Observed).

Note that C(A)>C(B) indicates that Forecast A has better performance than Forecast B.

The following table shows C(HiDyn), the correlation coefficient of anomalies of the 7-month-lead HiDyn-Model forecast and C(Persistence), the 7-month-lead Persistence forecast, where the correlations are calculated separately for two 15-year-long periods, 1987-2001 and 1992-2006.

Correlation coefficients for other ENSO forecast models in the Nino 3.4 Region can be found in van Oldenborgh et al. (2005). The correlation coefficients listed in the table above correspond to a "+6" monthly Nino-3.4 forecast in the terminology of van Oldenborgh et al. (2005). However, they do not go beyond "+5" monthly Nino 3.4 forecasts, so our results and theirs cannot be directly compared.

Performance measures based on correlation coefficients have the disadvantage that a forecast could be biased but still have a maximum skill of 1 (a correlation coefficient does not distinguish between forecast=observed and forecast=0.5*observed +1°C). The HiDyn Model is built to forecast well at the 1° × 1° pixel scale in the whole Tropical Pacific Ocean, not just at a regional scale on Nino 3.4. Another skill measure that might be tried is whether the warm/normal/cold regimes of the forecast are correct (These regimes are defined in Berliner et al., 2000). The HiDyn Model should have very high performance in this regard, because of the way it is built. 

Reference:

L.M. Berliner, C.K. Wikle, and N. Cressie, 2000. Long-Lead Prediction of Pacific SSTs via Bayesian Dynamic Modeling. Journal of Climate, 13, 3953-3968.

G.J. van Oldenborgh, M.A. Balmaseda, L. Ferranti, T.N. Stockdale, and D.L.T. Anderson, 2005. Did the ECMWF Seasonal Forecast Model Outperform Statistical ENSO Forecast Models over the Last 15 Years? Journal of Climate, 18, 3240-3249.