Forecasting in EVA

[americast]

Implemented standalone forecasting in EVA (using statsforecast package). You can run it via the following commands:

DROP TABLE IF EXISTS AirData;

CREATE TABLE AirData (
    unique_id TEXT(30),
    ds TEXT(30),
    y INTEGER);

LOAD CSV 'data/forecasting/air-passengers.csv' INTO AirData;

DROP UDF IF EXISTS Forecast;

CREATE UDF Forecast
FROM (SELECT unique_id, ds, y FROM AirData)
TYPE Forecasting
'predict' 'y';

SELECT Forecast(12) FROM AirData;

Here Forecast(12) signifies a horizon length of 12.

Thanks!

[xzdandy]

Implemented standalone forecasting in EVA (using statsforecast package). You can run it via the following commands:

DROP TABLE IF EXISTS AirData;

CREATE TABLE AirData (
    unique_id TEXT(30),
    ds TEXT(30),
    y INTEGER);

LOAD CSV 'data/forecasting/air-passengers.csv' INTO AirData;

DROP UDF IF EXISTS Forecast;

CREATE UDF Forecast IMPL 'evadb/udfs/forecast.py';

SELECT Forecast(unique_id, ds, y) FROM AirData;

I plan to add more features to this. Tests and documentation are still pending.

Thanks!

Hi @americast, the design looks great. Is the idea that training will be implicit here? When the user runs SELECT Forecast(unique_id, ds, y) FROM AirData;, EvaDB will train and forecast underlying together. The rationale is that the trained forecast model can only apply to the same data source.

[americast]

Thanks @xzdandy for your review. As of now, yes the training is implicit. Since statsforecast trains really fast, that should be fine. However, if we were to incorporate DL-based forecasting into this, we might want to train explicitly in the background.

[jyotigoyal09]

Hi @americast, I see you are using Forecast(unique_id, ds, y) to forecast for a single time series data. I was wondering if you have a functionality to forecast for panel data. If yes, how would you make final model selection if groups in the panel do not have same seasonality/trend pattern.

[gaurav274]

Hi @americast, I see you are using Forecast(unique_id, ds, y) to forecast for a single time series data. I was wondering if you have a functionality to forecast for panel data. If yes, how would you make final model selection if groups in the panel do not have same seasonality/trend pattern.

Hi Jyoti, Thanks for showing interest in evadb. We are in the early phases of adding forecasting. @americast @xzdandy Thoughts?

[americast]

Hi @americast, I see you are using Forecast(unique_id, ds, y) to forecast for a single time series data. I was wondering if you have a functionality to forecast for panel data. If yes, how would you make final model selection if groups in the panel do not have same seasonality/trend pattern.

Hi @jyotigoyal09. Thanks for your interest! Having the functionality to forecast for panel data could be a very useful functionality. In case of miltivariate forecasting, we can find the lowest common seasonality and use the same. However, we could also focus our attention on various deep learning models that do not have the requirement for specifying a seasonality upfront, such as transformer-based models.

It would be great if we could discuss in detail about this in an issue.

[americast]

@xzdandy I have pushed some commits and now the training is more like the Ludwig-style (#935). The training occurs when the UDF is created and the trained model is stored using a unique model file name generated from the data. Now, whenever the UDF is called, the user only needs to specify the horizon. I have updated the commands in #969 (comment) accordingly.

[xzdandy]

Thanks for the contribution! The implementation looks good to me at high level.

We can add

• Add a long integration test cases to verify it worked end-to-end.
• Clean up the code (remove redundant print, redundant import)

[americast]

Thanks for the contribution! The implementation looks good to me at high level.

We can add

• Add a long integration test cases to verify it worked end-to-end.
• Clean up the code (remove redundant print, redundant import)

@xzdandy Thanks. I have added the test and the docs.

[xzdandy]

Thanks for adding the tests and documentations. They look great.

Minor improvements.

1. Fix the linter: bash script/test/test.sh -m LINTER
2. Add statsforecast dependency in setup.py
3. Install the statsforecast dependency in .cricle/config.yml, so the test will be run in long integration tests.

[americast]

Thanks for adding the tests and documentations. They look great.

Minor improvements.

1. Fix the linter: bash script/test/test.sh -m LINTER
2. Add statsforecast dependency in setup.py
3. Install the statsforecast dependency in .cricle/config.yml, so the test will be run in long integration tests.

Thanks. I have updated them. Please let me know if this looks all good.

[xzdandy]

Great work! Verified the long integration test works on python3.10.

[americast]

Great work! Verified the long integration test works on python3.10.

Awesome, thanks!