Releases | Open Source Science

https://github.com/aouyang1/go-forecaster - v0.3.11

optimization on string building to reduce run time at inference (10x inference improvement)
prevent unnecessary buffering of time slice for rectangular window
add interface for event series

- Go
Published by aouyang1 6 months ago

https://github.com/aouyang1/go-forecaster - v0.3.10

performance improvements in reducing hopping around in memory to find the next feature slice to train on
added minimum uncertainty value to control minimum band width in cases that variance is non-existent and we still want some forgiveness
refactoring

- Go
Published by aouyang1 6 months ago

https://github.com/aouyang1/go-forecaster - v0.3.9

invert residual to represent y - yhat

- Go
Published by aouyang1 10 months ago

https://github.com/aouyang1/go-forecaster - v0.3.8

Refactored intercept to be a growth feature Helper method for custom seasonality

- Go
Published by aouyang1 11 months ago

https://github.com/aouyang1/go-forecaster - v0.3.7

add use_log in forecast options bug fix for estimating frequency with same counts better error handling

- Go
Published by aouyang1 11 months ago

https://github.com/aouyang1/go-forecaster - v0.3.6

Added Tukey window for smoother transitions into and out of an event window Refactored tests

- Go
Published by aouyang1 about 1 year ago

https://github.com/aouyang1/go-forecaster - v0.3.5

Removed unrolling of float operations to evaluate performance

- Go
Published by aouyang1 about 1 year ago

https://github.com/aouyang1/go-forecaster - v0.3.4

updated auto lasso to precompute data structures and reuse instead of redoing per lambda

- Go
Published by aouyang1 about 1 year ago

https://github.com/aouyang1/go-forecaster - v0.3.3

Improve lasso fit performance (~25-40%) by unrolling slice float operations by a factor of 4

- Go
Published by aouyang1 about 1 year ago

https://github.com/aouyang1/go-forecaster - v0.3.2

bug fix support for weekends and events on edge of training window with non-rectangle window function

- Go
Published by aouyang1 about 1 year ago

https://github.com/aouyang1/go-forecaster - v0.3.1

added auto lasso to pick best lambda regularization parameter and integrated with forecasting
refactored seasonality to support custom seasonality and removed daily and weekly params
include model events in table print
only check r2 score if positive or zero
update github actions to integrade with codecov.io
refactored to satisfy some sonarqube maintainability issues
moved forecast options into its own package
- moved much of the feature generation logic tied to each option
event end time is now exclusive instead of inclusive

- Go
Published by aouyang1 about 1 year ago

https://github.com/aouyang1/go-forecaster - v0.3.0

moved changepoint and event under forecast package
added comments and tests for feature package and general test coverage

- Go
Published by aouyang1 about 1 year ago

https://github.com/aouyang1/go-forecaster - v0.2.11

added weekend buffer option to account for window methods on weekends
refactored time and signal generation into timedataset package
added complete tests for generating time/seasonality features
removing zero only features from training
use filtered changepoints when generating changepoint features
added more test coverage to forecast utils
added tests for event creation
refactored events to be a separate feature type
include event masks as part of the model

- Go
Published by aouyang1 about 1 year ago

https://github.com/aouyang1/go-forecaster - v0.2.10

Added custom window function for events and weekend modelling. Defaults to rectangular windowing, but may introduce some discontinuities as the weekend and event boundaries. Recommending to use Hann or FlatTop to start.

- Go
Published by aouyang1 about 1 year ago

https://github.com/aouyang1/go-forecaster - v0.2.9

Made both the iteration and tolerance Lasso parameters configurable for forecasting. This can be set independently for the series and uncertainty fit. These both have large impacts on fitting times where uncertainty fit may want a lower iteration count or more flexible tolerance e.g. series has 500 iterations with 1e-3 tolerance and uncertainty with 250 iterations with 1e-2 tolerance

- Go
Published by aouyang1 about 1 year ago

https://github.com/aouyang1/go-forecaster - v0.2.8

Added event seasonality adjustment. Can add arbitrary time ranges and helper methods to cover Thanksgiving and Christmas with buffer time before and after holidays.

- Go
Published by aouyang1 about 1 year ago

https://github.com/aouyang1/go-forecaster - v0.2.7

Added seasonality changepoints for modeling weekend changes. Options provide ability to change the offset of when to begin modeling the weekend.

- Go
Published by aouyang1 about 1 year ago

https://github.com/aouyang1/go-forecaster - v0.2.6

add option to account for behavioral time shifts based on daylight savings time (DST). This shifts the behavior by the specified time zone for timestamps that are in DST. In the presence of multiple time zones specified, the average time offset for a given time is used.

Example: Europe/London + America/Los_Angeles timezones * Pre 2024-10-27 - offset will be 1hr since both Europe and America is in DST * 2024-11-01 to 2024-11-03 will offset by 30min since Europe is out of DST and America is in DST * Post 2024-11-03 the offset be 0.