4.7 Station Analyses


Eight analyses have been included to generate and analyse forecasts of individual predictand data. Analyses include:

  1. Probability distributions: Forecasts are presented in the form of probability distributions for both stratification and discriminant analysis predictive systems, accompanied by distributions of climatology. Stratification-based outputs are presented as multiple distributions representing separate stratifications with current conditions in bold. Discriminant analysis based outputs are presented as single distributions generated from averaging upper and lower envelopes from multiple two-category discriminant analysis iterations (see Chiew and Siriwardena, 2005). The user can interact with the plotted distributions to reveal interpolated values. Historical events (dates) can also be superimposed onto the curves.
  2. Probability pie charts: Tercile and above/below median outlooks are presented here in the form of pie charts (sometimes called ‘chocolate wheels’). This represents the simplest way to disseminate forecast information to end-users. Training size, LEPS skill score and percent consistent values accompany each pie chart.
  3. Box plot analysis: Box plots representing different stratifications are presented here to help visualize and compare the variations between stratified samples. This is particularly useful when differentiating the impacts of EL Nino and La Nina climate extremes.
  4. Sampling regression analysis: Scatter-plots of predictor versus predictand analogues are presented to show their correlation. Equations are generated which can be used for regression based forecasting. Multiple scatter-plots are generated when using discriminant analysis based predictive systems containing multiple predictor elements. Points can be hot-tracked to show which event they relate to.
  5. Skill score analysis: A table or “skill map” of skill scores (LEPS, Percent consistent, or ROC) is presented for a range of inter-annual forecast periods and lead-times. The map represents forecast skill results of 108 separate “cross-validated hindcast” analyses (12 periods by 9 lead times, by default). The forecast periods are represented on the x-axis, with the lead-times on the y-axis. The skill score results are assigned colours relative to the magnitude of each score: a blue square denotes forecasting skill greater than climatology (chance); a red square denotes forecasting skill worse than climatology; while a white square denotes skill the same as climatology. A circle within the skill map represents the current period-setter conditions. Users can click on any square within the skill map to automatically synchronise the period-setter for those conditions.
  6. Hindcast performance calendar: A table of all hindcast results over the defined testing period (same as the training period by default) for each successive month is presented here. Results are colour-coded according to a variety of user-defined schemes; discrete colours based on hindcasts being ‘consistent’, ‘near-consistent’, or ‘inconsistent’; graduated colours based of the previous scheme, but with colour-weightings representing forecast-strength (largest probability size); and graduated colours based on LEPS skill scores. This table is also interactive allowing the user to synchronies the period setter with individual cells. All of the graduated colour schemes provide visual information on how good a particular forecast was, with very good forecasts showing in strong blue colours, and very poor forecasts showing in strong red colours. Hindcasts similar to climatology (‘unskilled’) are represented in close-to-white colours regardless of whether they were consistent or inconsistent.
  7. Seasonal hindcast analysis: A portion of the output from the previous analysis has been extracted and displayed here representing the hindcast performance for the current predictand period. A timeline of hindcast years is split into vertical tercile (or above/below median) groupings with observed (O) and predicted (P) categories highlighted. The colouring scheme from the previous analysis has been adopted. Also, a detailed hindcast evaluation is available in the report view with numerical outputs for validating and investigation individual hindcasts.
  8. Historical analogue analysis: This analysis presents an alternative view of the seasonal hindcast analysis in bar chart form, with individual bars representing predictand analogues totals (or averages) colour-coded according to the previously defined hindcast colour schemes. For stratification based predictive systems, bars can also be colour-tagged based on the corresponding stratification phase for each year.


Figure 10: Examples of station analyses including (a) stratification probability distributions; (b) tercile probability pie charts; (c) box plots of stratifications; (d) predictor/predictand regression analysis; (e) tercile LEPS skill score map; (f) hindcast performance calendar; (g) seasonal hindcast analysis; and (h) historical analogue analysis.Figure 10: Examples of station analyses including (a) stratification probability distributions; (b) tercile probability pie charts; (c) box plots of stratifications; (d) predictor/predictand regression analysis; (e) tercile LEPS skill score map; (f) hindcast performance calendar; (g) seasonal hindcast analysis; and (h) historical analogue analysis.