Come to this talk if you want to learn a few basic techniques for putting numerical data in context. If you've ever predicted anything, or tried to work out whether some number was "good enough", you'll probably get something out of this presentation. All techniques and tools demonstrated using Python.
Every day, decisions both big and small are made on the basis of the information published by the Bureau of Meteorology. These include simple decisions such as taking an umbrella or planning a barbecue. Our forecasts also inform Australia's emergency services on where extreme weather events may have occurred, to help with planning and preparation.
Understanding and communicating our strengths and weaknesses is very important, both as an organisation and also internally within the Environment and Research division. This presentation will focus on the statistical methods and systems used to evaluate the objective, scientific performance of our forecast systems. The name for this area of study is "Verification". While the concepts have come from the research environment, they are widely applicable and can help anyone who is assessing the performance of any system.
This presentation will include: -- An overview of the major ideas of verification -- How to create a 'skill score' -- The application of these concepts to thunderstorm forecasting -- How to use Python tools for verification analyses -- Tips on how to apply these ideas easily in other contexts
Obtaining relevant thunderstorm observational data can be particularly challenging, particularly pertaining to severe and damaging aspects: lightning, hail, heavy rain and very strong wind gusts. In order achieve a stronger footing, some new methods of analysis are under development. It is necessary to establish the scientific validity of the verification metrics at the same time as constructing the systems to support the data analysis.
A prototype web-based tool written in Python (and under active development by the presenter) will be demonstrated. This tool can run locally to provide an enhanced lab environment for assessing case study data, or be set up as a server for continuous monitoring and reporting.
No pre-existing knowledge of Python or statistics is assumed. The talk will include several technical aspects, such as working at different computing scales, usability and user experience, working with statistical algorithms, data visualisation for both web and journal publications, and the architectural challenges of a complex application.