Climbing Ratings 1.2.0

Changes since 1.1.0:

• Switched to using log-normally distributed priors for the "gamma" ratings, instead of gamma-distributed priors.
• In the routes.csv input to 02-run_estimation.py, rename grade to gamma. This means inputs generated with previous code will need need to be regenerated (or manually fixed).
• Drop climbers with no "non-clean" ascents in pre-processing.
• Stop classifying "tick" as clean.
• Some regression tests for top-level scripts.

Climbing Ratings 1.1.0

Changes since 1.0.0:

• Support for parsing data from theCrag:
  • parse logbook exports
  • parse JSON responses from the API
• Execution speed:
  • detect early convergence using log-loss function
  • optimize some hot parts of the estimation
• Additional statistics:
  • log-loss, Cohen's kappa, route ratings variance
  • k-fold cross validation
  • hyperparameter search

Known issues

• There are still some constants that aren't being optimized by the hyperparameter search
• There are some source of bias in real data that might be rectifiable with additional heuristics
• The top-level scripts lack test coverage

climbing_ratings 1.0.0

Initial release

The rating outputs appear sensible and the code is fast enough to fit global databases of ascents in minutes.

Known issues

• The Python package isn't in a state to be registered - setup.py doesn't have all the registry fields set
• The Cython is getting compiled with clang-specific CFLAGS and some aggressive options (that can result in undefined behaviour if there's a bug!)
• There's no Python functionality for evaluating the model's log-likelihood function.
• There's no robust analysis of the model's suitability for real-world climbing data, or comparisons to other models.