Probabilistic Forecasting Exploratory Development

Strawman Forecast Process

After being introduced to the GSD prototype tools, lively discussions and feedback led forecasters to a strawman forecast process. The concept is to incrementally modify current practices for producing deterministic forecast products (i.e., those currently posted to NDFD), by providing options to invoke any of several types of statistical techniques and/or guidance to help the forecasters wrap uncertainty components around the official forecast in an efficient manner that leaves room for forecaster input or quality control. This implies that the current NDFD deterministic grid product, temperature for example, is to be interpreted as the median of the forecast PDF.

In experimenting with the prototype software, the high level steps described below are suggested as a starting point.

After reading through the high level steps, we suggest trying out the Forecast Process Job Sheet which steps you through one possible end-to-end forecast scenario using the prototype software and canned data set. This will give you a basic understanding of the process and act as a launching point for exploring all the various tools in the prototype software. More in-depth reference material and cheat sheets for all capabilities are accessible in the Forecast Process pull-down menus.

We then welcome any and all feedback/discussion which can be posted on the Probabilistic Forecasting listserver. (For instructions on joining, please contact us). Thanks for your participation!

High-Level Forecast Process Steps

  • Visualize and Diagnose Ensemble Data: Forecasters study guidance, both deterministic and ensemble through various visualization techniques in ALPS and GFE, i.e. Ensemble Statistics (Mean, Median, and Std Dev Temperature Displays), Ensemble Histograms, and Ensemble Tables, and in 3D Visualization.
  • Weight the Ensemble Members: Optionally, forecasters may want to weight the ensemble members to be used in producing the probabilistic grids. Forecasters can locally set up their weights database and edit ensemble weights using the Ensemble Weights Tool and the GFE Ensemble Relative Frequency (ERF) Editor.
  • Produce Deterministic Grids: Forecasters produce deterministic grids as they do today. This will represent the 50th percentile in the probabilistic distribution. Forecasters can choose an ensemble (or blend of ensembles) from which to initialize the grids.
  • Produce Probabilistic Grids: Forecasters produce probabilistic grids by choosing from various methods
    • "GenerateProbGrids" Smart Tool will create 10 & 90 grids from the 50 grid given a set of possibly weighted ensemble members.
    • Future work:
      • Precip: Amburn method for estimating PQPF from PoP and QPF
      • Temp: Krzysztofowicz method for estimating spread based on the deterministic temperature and NDFD verification statistics.
      • An estimation of today's forecast spread in T based on yesterday's observed spread in temperatures.
      • EKDMOS approach for first guesses for spread
      • Blend QPF coverage from NAM and areal mean from SREF
      • NCEP methods
    In some cases, they will choose guidance source(s) upon which to base the distribution, possibly weighting the guidance members, and choosing whether to generate simply endpoints or an entire spread of values in the distribution.
  • Generate Products: For ideas on products from the 10-50-90 grids, we'd like your feedback! To get an idea of some possible probabilistic phrasing, run the experimental GFE Probabilistic Phrasing Text Product which is generated directly from the ensemble members (weighted or unweighted). Also, check out the mock-up for Probabilistic Visualization. Work is underway to generate this graphic directly from the 10-50-90 grids.
Graphic representation of Forecast Process