Timeseries Forecasting

Timeseries forecasting has multiple applications like predicting inventory, trends in stock prices, housing trends, weather and sales predictions to name a few. Users have a variety of timeseries predictive models to choose from. Which type of model works best depends on the type of task and data that you are using. Tiyaro is making a large number of timeseries models available that you can easily and quickly train, compare and choose the model that works best for you without having to write a single line of code. There are 2 types of predictive models that are currently supported.

  • Univariate Forecasting

  • Univariate Forecasting with exogeneous variables

Univaraite Forecasting

The following univariate models are available on Tiyaro

  • Prophet

  • AutoARIMA

  • LightGBM

  • RandomForest

  • XGBoost


  • CatBoost

Univariate Forecasting with exogeneous variables

The following univariate models with exogeneous variables are available on Tiyaro

  • LightGBM

  • RandomForest

  • XGBoost

  • CatBoost


Timeseries Forecasting models are different from the other AI/ML models being served by Tiyaro. Forecasting is very specific to a user and their historical data. Hence the typical workflow for using a Timeseries model involves training the various models for your dataset and then using the trained models for inference. The details of these steps are covered below.

In order for you to customize or train the models all you need is the historical data that will be used to train the models. The entire process of training is simplified by the Tiyaro EasyTrain system. We built the EasyTrain system so that users can train all the algorithms in a single step without writing a single line of code. EasyTrain also uses intelligent defaults for the various hyperparameters requried by the models so the burden of training the models is shifted from the user to the EasyTrain system.

How to train your custom timeseries predictive models

  1. Find the Timeseries base model.

  2. Train or Customize the base model with your training data

Find the Timeseries base model

Simply search for ‘timeseries’ or the name of any of the models from above .e.g. prophet, autorarima in the Model Explorer.


Select the Time Series Forecasting model. It has type of timeseries-base class. This is the base class form which you can train your custom models. Selecting this model will open the model card as shown below.


Note the Run EasyTrain button on the model card

Train/Customize the base model with your training data

Click on the Run EasyTrain button to start the training. It will walk you through the steps of training the timeseries model.

First step, name your training project.


Second step, provide your training data.


Format of the expected training dataset

The training data needs to be packaged as a zip file. Here is a sample dataset that you can download to inspect its format and structure. Lets look at the structure of this sample dataset to understand the expected format.

$ wget -q https://public-model-demo.s3.us-west-2.amazonaws.com/timeseries-sample-dataset.zip

$ ls

$ unzip timeseries-sample-dataset.zip
Archive:  timeseries-sample-dataset.zip
  inflating: temperature.csv

$ ls
temperature.csv  timeseries-sample-dataset.zip

$ head temperature.csv

As you can see above, the zip file timeseries-sample-dataset.zip contains a single csv file called temperature.csv. The csv file has many columns. Where

  • date - is the datetime column

  • meantemp - is the target column i.e. the value that is going to be predicted

  • wind, speed, meanpressure - are the exogeneous input columns. Note these are optional columns

Last step, Finish.


Once you click Submit the training job is kicked off. It typically takes 10-40 minutes to train a model depending on the size of your training datasaet and the number of algorithms/models that you are training.

Training jobs

All your training jobs are listed under the Train menu item in the UI as shown below.


Clicking on one of the jobs will open up the training card. On the training card the Config subtab shows the details of inputs used to train the job as shown below.


The Results tab shows the custom trained models as shown below


Clicking on them will open the Model card for those models as shown below. Like every model card it has all the information including the API endpoint, swagger spec for you to use that model.


In case you are wondering what about Univariate vs Univaraite with exogeneous inputs. The EasyTrain system takes care of it for your automatically, as explained in the note below.


If you provide the optional ‘Input Columns’ in the training input then the EasyTrain system will train Univariate with exogenous input models else it trains the Univariate models


We saw the process of training your custom timeseries forecasting models to find the trained models in the Trainn tab to going to the model card of the trained models where you find all the information to use those trained models in your application. You can also search for your models in the Model Explorer.