 This is a demo on AutoAI time series. We show a new capability of AutoAI on time series forecasting. Time series has a very wide range of use cases in practice, such as financial services, IoT, and retail. It is critical to provide fast and accurate results with minimum requirement on human expertise. AutoAI time series is a fully automated tool which does not require a line of code from the user with few clicks, it generates a rich set of optimized statistical and machine learning pipelines. Now let's see how it works. First, user needs to upload a data file with one click, upload a file, select data. Now the data has been loaded. This is a unique varied time series on daily minimum temperature. Here you can see some basic information including the number of columns and the size. The second click is on the column you want to forecast. Here we select temperature. You can go to the advanced setting. If you want to specify additional parameters, this is optional for advanced users. Prediction type, we select time series forecasting. Look back window. It is how many past value you think is relevant for your forecasting. You can choose the building algorithm to select the look back window automatically. Forecasting window is on how many steps ahead you want to forecast. The default value is one. You could also choose the optimized matrix for your pipeline selection. Here we have root mean square error, mean absolute error, and an SMAP. Algorithms to test is which algorithms you want your pipeline to use. As you can see, we include a wide range of statistical and machine learning algorithms. If you do not want certain algorithms, you just click and to exclude it. You could also choose the number of pipelines to output. Here we choose five. We save settings and start the experiment. Now the experiment has started, gave the training data. The tool begins to search for pipelines demonstrated by this dynamic visualization. The progress report shows the real-time activities happening at the back end. Starting from the algorithm, the pipelines are experimenting with different combinations of feature transformers and estimators. For example, pipeline 1 used window-in transformer and imputation. Pipeline candidate board shows the real-time ranking of the partially fitted pipelines. It will keep flashing until the end the top pipeline are selected. We can swap the view to progress map. This flow chart shows you which stage the current backend is running. The flashing dot shows you the model selection. The solid dots indicates the pipeline has been finished. As you can see, for example, pipeline P7 is a hot winters, has been finished. This is the RMSE at the end, 3.824. However, this ranking is not final yet. And the progress report shows currently it's evaluating pipeline P10. Now the experiment has completed. As we indicated from the beginning, we'd like to have 5 top-picked pipelines at the end. So you can see, here are the 5 top-picked pipelines and there are 5 discarded pipelines. Now the pipeline candidate board has switched to the leader board to show you the final ranking of the pipelines. The one with stars is the top-picked pipeline P7. It's used hot winters, additive. If we click onto the pipeline, it shows you some information of this pipeline. Notice that the Type Series tool provides a user with 2 different scoring schema. One is on the held-out data, one is using the back test. You can switch between these 2 different schema. On the top, notice that the IMCE changed when you select these 2 options. 3.019 is the back test RMSC for pipeline 7 and 1.897 is the hold-out RMSC for pipeline 7. For each of the scoring schema, we provide the user with 4 different evaluation matrix. And we also have a sequence chart to show you the comparison between the actual time series value versus the predicted ones. And this is the option for all the pipelines on leader board. Moreover, you can drill down into the pipeline to get more informations. For example, pipeline 3 is currently ranked as number 2. In the pipeline viewer, we provide an overview on the different parameters for fitting pipeline 3. For example, the lookback window, the number of back tests we used and the prediction horizon right now is 1. And here we also provide you the scores on the back test and the hold-out. In addition, the performance over time shows you the actual values versus predicted values and the layout of multiple back test data allocation plus the hold-out section. For the stability, it shows you the arrow in each back test and the hold-out section and you can choose different matrix. In this particular example, you can see in back test allocation 0. All three matrix shows particularly high arrow. So the stability will provide you with a view on how the pipeline is performing in different allocation of data sections. If you decide to use this particular pipeline for your forecasting job, later you can save it either as a model or as a notebook. If we go back to the main page you could drill them into the details for other pipelines on the leaderboard. After the pipelines are saved then it's ready to forecast. If you want to try another data set you can reconfigure the experiment. The daily minimum temperature is a univerity time series. Next, I will demonstrate briefly on a multiverse time series. On this particular NN5 data set, as you can see the data has been loaded as before. However, now in this data set we have 112 columns. And among these columns you can make multiple selections of the columns which you want to produce forecasting. The rest of the experiment setting is the same as the univerity case which I will not elaborate. With one click of the button you can start running another experiment. This is the end of the demo and thank you very much for watching.