July 25, 2022

Miscellaneous improvements

• Foundational data: Added a new source from Factset for daily FX rates covering 51 currencies, which is broader than what was previously available through Quandl. This is now used by default in the fx() and signal.convert_currency() functions. You may still use the Quandl source by specifying the source parameter (reference).
• DSL: Add signal.change() function (reference) to calculate absolute change in the signal over a specified date offset. This generalizes the existing signal.yoy(), signal.qoq(), and related functions.
• DSL: Add mae(...) function (reference) to calculate the mean absolute error between a target and predicted signal. This is useful for users who want to export model errors, as the backtest absolute error reported in the prediction model user interface is equivalent to mae(target_signal, backtests(model_number, backtest_number)).
• DSL: Add support for delta transform in r2_score, wape, regression_weight, correlation and centered_correlation functions (reference) .

Bug fixes

• Classification models: remove NaNs as part of precomputing features, which was causing model failures with walk-forward backtesting in some cases.