System and method for determining a transit prediction model

1. A method, comprising:

a. training a set of models using supervised learning, wherein each model is associated with a sliding training window, and wherein each model is trained on a different set of historic transit data selected based on the respective sliding training window;

b. determining an actual transit time for each package in a set of packages delivered within an evaluation period;

c. for each model in the set of models:

determining a predicted transit time for each package in the set of packages using the trained model;

determining individual evaluation metrics for each day of an evaluation period based on the respective predicted transit time and actual transit time for the set of packages; and

determining an overall evaluation metric by aggregating the respective individual evaluation metrics;

d. selecting a model from the set of models based on the respective overall evaluation metric for each of the set of models; and

e. using the selected model, predicting a future transit time for a target package, wherein the target package is associated with a shipment created within a prediction period.

2. The method of claim 1 , further comprising repeating a)-e) for a successive prediction period.

3. The method of claim 1 , wherein the evaluation period is redetermined for the prediction period.

4. The method of claim 1 , wherein the sliding training window comprises a set of dates relative to a date associated with a package, wherein at least two packages in the set of packages are associated with different dates, and the sliding training window encompasses a first set of dates for the first package and slides to encompass a second set of dates for the second package.

5. The method of claim 4 , wherein the date associated with a package is based on the respective shipment creation date.

6. The method of claim 4 , wherein the set of dates comprises nonconsecutive dates.

7. The method of claim 1 , wherein determining the predicted transit time comprises determining a minimum transit time for a predetermined percentile of packages in the historic transit data to be delivered.

8. The method of claim 1 , further comprising:

receiving, from a client computing system, a shipment request for the target package;

generating a response to the shipment request, wherein the response comprises the future transit time for delivering the target package; and

providing the response to the client computing system.

9. The method of claim 1 , wherein the historic transit data is associated with at least one of: a shipping carrier, a shipping carrier service, or a shipping lane.

10. The method of claim 1 , wherein the historic transit data comprises bi-directional transit times.

11. The method of claim 1 , wherein predicting the future transit time for the target package comprises:

determining a set of historic transit data based on the sliding training window, associated with the selected model, that is determined based on a date associated with the target package; and

retraining the selected model on the set of historic transit data, wherein the future transit time is predicted using the retrained model.

12. A method, comprising:

a. training a set of models, wherein each model is a machine learning model trained on a different set of historic transit data;

b. determining actual transit times for a set of packages;

c. for each model:

predicting a transit time for each package in the set of packages with the trained model;

determining individual evaluation metrics for each day of an evaluation period based on the predicted and actual transit times for the set of packages; and

determining an overall evaluation metric by aggregating the individual evaluation metrics;

d. selecting a model from the set of models based on the respective overall evaluation metric for each of the set of models;

e. using the selected model, predicting a new transit time for a package, wherein the package is associated with a shipment created within a prediction period; and

f. repeating a)-e) for a successive prediction period, wherein the set of models are retrained for the successive prediction period.

13. The method of claim 12 , wherein the different set of historic transit data for each model is determined based on a sliding training window associated with the model, wherein the sliding training window slides with a reference date.

14. The method of claim 13 , wherein the reference date is based on the prediction period.

15. The method of claim 12 , wherein the set of packages are delivered within the evaluation period.

16. The method of claim 15 , wherein d) is performed using a selection rule, wherein the evaluation period is associated with the selection rule.

17. The method of claim 16 , further comprising selecting the selection rule from a set of rules, wherein selecting the selection rule comprises:

for each rule of the set, selecting a model from the set of models based on the rule, wherein e) is performed using each selected model;

determining a prediction metric for each rule based on the new transit times predicted by the model selected by the respective rule; and

selecting the selection rule from the set of rules based on the respective prediction metrics.

18. The method of claim 12 , wherein selecting the model is further based on a user input.