Using Consistency-Based Forecasting to Extend Numerical Model Skill into Subseasonal Lead Times

Accurate subseasonal forecasts (3–6 weeks) of weather phenomena are of great interest to stakeholders and emergency managers. For example, there is a keen interest along the U.S. West Coast to have as much forewarning as possible for periods of anomalous atmospheric river activity, as they can lead to periods of drought or flooding depending on their absence or presence. Unfortunately, state-of-the-art numerical models only have skillful lead times of 2–3 weeks. Thus, there is a need to develop forecasting methods that can extend skillful forecast lead times.

Here, we hypothesize that the skill scores realized by numerical models can be improved at subseasonal lead times through a “consistency-based” forecasting method. We define a consistency-based forecast as one that harnesses information from both the current model run and all prior model runs that see the same forecast valid date rather than just the current model run. To test this hypothesis, we develop consistency-based, deterministic forecasts from the ensemble suites of NCEP and ECMWF reforecast data acquired from the S2S Project Database. The consistency-based forecasts are verified against ERA-Interim reanalysis data and compared to “current-based,” deterministic forecasts that are based solely on the current model run.

We make four sets of deterministic forecasts (F1 through F4). Specifically, we forecast for the 500-hPa geopotential height anomaly of a key region near the West Coast that is known to modulate atmospheric river activity in California. In forecasts F1 and F2, we base our prediction solely on the current model run’s ensemble suite. We label these our current-based forecasts. For F1, we use the ensemble mean to predict the sign of the anomaly. For F2, we use the individual ensemble members and only make predictions when a certain threshold of members agree on the sign of the anomaly. In forecasts F3 and F4, we base our prediction not only on the current model run’s ensemble suite, but also on the suites of all model runs prior to the current model run that see the same forecast valid date. We label these our consistency-based forecasts. For F3, we use the ensemble means from all suites and only make predictions when a certain threshold of ensemble means agree on the sign of the anomaly. For F4, we use the individual ensemble members from all suites and only make predictions when a certain threshold of ensemble members agree on the sign of the anomaly.

All predictions are evaluated using the Heidke Skill Score (HSS). It is found that the consistency-based forecasts (F3 and F4) have higher HSSs at longer lead times than their current-based counterparts (F1 and F2). Moreover, the HSSs for F3 and F4 increase as the threshold is increased, but this comes at the expense of a decreased forecast sample size. Comparing F3 to F4, F4 has modestly higher HSSs using not only fewer ensemble suites but also lower thresholds. The implication of this result is that F4 performs the best at subseasonal lead times and has a large forecast sample size. Our results demonstrate that consistency-based forecasting can be used to improve subseasonal forecasts of 500-hPa geopotential height anomalies off the U.S. West Coast. The results from this study are promising, and we propose that consistency-based forecasting should be tested on other variables for other regions in order to improve subseasonal forecasts of important weather phenomena.