Predict future data

Introduction

Once we have trained a forecaster and validate its performance and accuracy, it’s time to predict future data. In this guidance, we demonstrate how to use forecaster to predict future data with and without ground truth in detail.

We will take TCNForecaster and nyc_taxi dataset as an example in this guide.

Setup

To begin our guide, you should prepare your environment and your dataset as introduced in former guides.

Forecaster preparation

Before prediction, a forecaster should be created and trained. The training process is introduced in the previous guidance Train forcaster on single node in detail, therefore we directly create and train a TCNForecaster based on the nyc taxi dataset.

Prediction

There are two application scenarios for prediction : prediction with ground truth to validate and pure prediction (without ground truth).

For different kinds of prediction, the process about how to prepare your dataset differs. We will introduce the examples for different scenarios separately.

1. Prediction with ground truth

When a forecaster is trained, we provide with predict to predict on test data. The data support following formats:

1. bigdl.chronos.data.TSDataset (recommended)

2. numpy ndarray

3. pytorch dataloader

4. xshard item

It’s always recommended to use TSDataset direcetly when possible.

The batch_size will not affect prediction result but will affect resources cost and the default value is 32. Besides, quantize can be set to True to use the quantized model to predict.

# get TSDataset for training and testing
tsdata_train, tsdata_test = get_data()
tsdata_train.roll(lookback=48, horizon=1)
# get a trained forecaster
forecaster = get_trained_forecaster(tsdata_train)

# if predict with ground truth, you should set parameter in roll like this:
tsdata_test.roll(lookback=48, horizon=1)
# prediction using forecaster
pred = forecaster.predict(tsdata_test)

📝Note

If you are using AutoformerForecaster, don’t forget to set time_enc=True in tsdataset.roll.

Congrats! So far you have already finished your first try of forecaster.predict successfully. But there is one more thing you need to do.

We should know that we scaled our datasets both for trainning and for predicting by using tsdataset.scale when we prepared the dataset before. So the result of prediction here (denoted by pred) is also scaled. To get the result in original scale, we have to use tsdataset.unscale_numpy to unscale the result.

The input of tsdataset.unscale_numpy is a numpy ndarray, which is actually the return value of forecaster.predict function. So what we need to do is just passing the result into the tsdataset.unscale_numpy function to get the unscaled result.

# Unscale the predict value to get the result in correct scale
pred_unscale = tsdata_test.unscale_numpy(data=pred)

2. Prediction without ground truth (pure prediction)

To make a pure prediction, the only change is to set is_predict=True in tsdataset.roll. This will enable you to predict target features of unobserved timestamps.

# if predict without ground truth, you should set parameter in roll like this:
tsdata_test.roll(lookback=48, horizon=1, is_predict=True)
# prediction using forecaster
pred = forecaster.predict(tsdata_test)

📝Note

If you are using AutoformerForecaster, don’t forget to set time_enc=True in tsdataset.roll.

Then how to get the unscaled result is the same as before.