Bad RMSE when predicting Price with Linear Regression

Question

Hi. I have a dataset of price data. It looks like this

| Price| Branch| ItemCode| Discount| DateTimeOfPrice|
| --- | --- | --- | --- | --- |
| 10| 002| 52345436| 0.33| 2022-03-24 14:00|

The dataset has about 1M records

I feature engineered it in the following way

| Price| Discount| ItemCode| Year| Month| Day| Hour| Branch1| Branch2| Branch3|
| --- | --- | --- | --- | --- |--- | --- | --- | --- | --- |
| 10 | 0.33| 52345436| 2022 | 03 |24 | 14 | 0 | 1 | 0 |

Each component of the DateTimeOfPrice got a separate column
We have 3 branches. To avoid the situation when algorithm may think that "branch" column is some kind of priority column, I created 3 new column (we have 3 branches). If the item belongs to branch2, the column will get the value 1, if not - it will be 0

I run Linear Learner, XGBoost build-in algorithms and also SageMaker AutoPilot.
In all cases I run , the best RMSE was 60 and prediction/ validation gives sometimes a result which is far from the actual value.
I tried also to run XGBoost from the notebook with the following parameters

```
hyperparams = {
    "max_depth": "7",
    "eta": "0.2",
    "gamma": "4",
    "min_child_weight": "6",
    "subsample": "0.7",
    "objective": "reg:squarederror",
    "num_round": "100",
    "eval_metric":"rmse",
    "verbosity": "2",
}
```

Still, the RMSE is arround 60.

Please advice what can be done to improve the mertic and predication

Answer

I suggest before you start to build your algorithm, do a data exploration. Does your data have a seasonality? Some items are just not seasonal.

Answer

I followed you suggestion and used Sagemaker Canvas

I modified the data structure in the following way

| ItemPrice | Branch| Discount| ItemCode| PriceDate |
| --- | --- | --- | --- | --- |
| Data | Data | Data | Data | Data |
| Data | Data | Data | Data | Data |

I choose ItemCode as "id" and "grouped" by "branch".
However the score of the prediction is very poor score 22%

According to the analisys the reason is because of the Discount column. So I removed it and run the process again. And the score was even lower :(

Answer

Since I see you have a timestamp field in your data, would it be fair to assume your use case is mainly aimed at forecasting *future* prices - rather than estimating missing historical prices at different points in time?

If so, plain tabular regression (Autopilot regression task type) is probably not a good way to tackle this problem as forecasting techniques would work better instead. You could instead explore:
- SageMaker Canvas, which offers a forecasting model (see the [docs here](https://docs.aws.amazon.com/sagemaker/latest/dg/canvas-time-series.html) to make sure your input timestamp is recognised so that Canvas shows you the forecasting option)
- Amazon Forecast, a dedicated [managed forecasting service](https://docs.aws.amazon.com/forecast/latest/dg/what-is-forecast.html) separate from SageMaker

Bad RMSE when predicting Price with Linear Regression

Relevanter Inhalt