Group 8 Application

Proposal

Poster

Application

Report

Application Web Address

The application has been uploaded onto the following URL for your usage/viewing:

https://timeseriesexplorer.shinyapps.io/timeseriesexplorer/

User Guide

Please refer to the below steps for how to use the system.

Step 1:

Step 1 - System interface after initial load

When the app loads, users can click on the "Browse..." button in the "Upload Data File" tab, to select a time-series data .csv file to upload.

Step 2:

Step 2 - Meta table and data preview

After the upload is complete, users can see several sections of information.

On the left-hand side panel, the table displays the various column names (Name), data type (Type), the count of missing records (Mis Value), as well as the position of the missing values (First Mis Value, Last Mis Value).

The main screen on the right, will show a preview of the uploaded data, complete with filter, search, paging and sorting functions.

Step 2.1 (optional):

Step 2.1 - Transposing data

• If the uploaded dataset does not have a datetime column for time-series analysis, users have an option to click on "Add Index", to generate a new column named "newIndex" that is essentially a running number index column. This will then be converted into a pseudo-datetime column for visualization later.
• Users can also click on the "Transpose The Uploaded Dataset" radio button to perform some data transformations from row to column. After selecting the "From:" and "To" field, enter in the newly renamed "Column Name of Category" and "Column Name of Accumulated Value", before clicking on the "Resume/Transpose Data" button. The transposed columns will appear at the far-right of the preview table.
• For any desire to revert any changes made, simply click on the "Resume to The Original Dataset" radio button, and click again on the "Resume/Transpose Data" button to revert all changes back to the original dataset.
• Users can also click on the "Download" button to download the transformed dataset onto their local drive.

Step 3:

Step 3 - Filtering columns and data rows

Click on the "Explanatory" tab, and there are many filter options to use at the left-hand panel. In the "Select Column:" section, choose the data columns for "Time-Series Value", "Time Index" and "Category" to dynamically load the type of attributes to filter.

The main screen will display the time-series chart for the main observation data selected, along with its decomposed Seasonal, Trend and Random components.

These charts are useful to understand the behaviour of the time-series data.

Users can then change the "Categorical Attribute", "Date start with no missing value" and "Period" fields to filter out the records.

For the "Customise Visualisation" section, there is also an option to denote whether the time-series data is "multiplicative" or "additive".

Seasoned analysts would know that the calculations done for these two types of time-series data are quite different.

The "Time Series Frequency" can also be set, with the default value of "12" to denote months in a year.

Step 3.1 (optional):

Step 3.1 - Filtering sub-graphs

Users can also filter out the sub-graphs by clicking on their respective checkboxes at the end.

Step 4:

Step 4 - Model selection based on accuracy measures

After making the selections from the left-hand panel, click on the "Forecasting" tab to allow the system to run a series of calculations to find the best parameters for both Exponential Smoothing and ARIMA models.

After giving it sometime, a table will be generated on the main screen. These are the best models chosen for their lowest AIC/BIC values and users can compare the various accuracy scores (based on train data only).

Users can then click on the various rows for closer scrutiny and click on the "Forecast Charts" tab to generate line charts with their respective forecasts.

The "Forecast Period" slider on the left hand panel allows users to select how far ahead the charts can forecast up to.

To save the model parameters, click on the "Download Stats" button. To save the charts, simply right-click and copy the charts to the clipboard or save manually.

Step 4.1 (optional):

Step 4.1 - Forecasted Time-series Charts

There is also an option for the user to upload "Holdout data" by clicking on the browse button under the slider.

Once uploaded, the charts will automatically reflect the additional actual data in the line charts, for comparison with the predicted values.

Click on the "Clear Holdout" button to clear the holdout line from the charts.

Model Results Interpretation

For the actual forecasting results, we make use of the ets() and auto.arima() functions from the powerful "forecast" package. From the forecasting metrics table, one can see the results denoted in a certain format.

ETS represents Error, Trend, Seasonality. The letters A, M, and N stand for 'Additive', 'Multiplicative', and 'None' respectively.

Alpha, Beta, Gamma, and Phi are parameters used in their respective ETS formulae calculations.

Some popular combinations of ETS models and their alternative names include:

ETS(A,N,N) = Simple Exponential Smoothing with Additive Errors

ETS(A,A,A) = Additive Holt-Winters' Method with Additive Errors

ETS(M,A,M) = Multiplicative Holt-Winters' with Multiplicative Errors

ETS(A,A,N) = Additive Errors, Additive Trend and No Seasonality = Holt's Linear Method with Additive Errors

ARIMA is a short acronym that stands for Autoregressive Integrated Moving Average.

Also from the forecast package, ARIMA has three parameters like the ets() function: ARIMA(p, d, q).

p stands for 'Number of Autoregressive Term' or AR term.

d stands for 'Number of Non-seasonal differences' or I term.

q stands for 'Number of Lagged Forecast Error Terms' or Moving Average/MA Term.

Some popular combinations of ARIMA models include:

ARIMA(0,1,0) with drift = Random walk with constant/growth

ARIMA(1,1,0) = Differenced first order AR model

ARIMA(0,1,1) = Simple Exponential Smoothing

ARIMA(0,2,1) or ARIMA(0,2,2) = Linear Exponential Smoothing

ARIMA(1,1,1) with drift = Mixed Model

ARIMA(1,1,2) with drift = Damped-trend Linear Exponential Smoothing

For results with more than 1 set of brackets: i.e. ARIMA(1,1,2)(2,0,0)[12]

First set of brackets: Order Parameters

Second set of brackets: Seasonal Order Parameters

Third set of brackets: Period Frequency (for e.g. 12 months)

Choosing the models require looking closely at the measures or scores of the models.

Commonly, one can easily sort the table according to AIC (Akaike information criterion) and BIC (Bayesian information criterion) values. These are goodness-of-fit values based on training data and the minimum scores depict the best models for the data. Other useful measures would include Mean Error (ME) and Root Mean Squared Error (RMSE).

Feedback

We appreciate your feedback, please send us any comments via the email addresses below:

guoteng.fam.2016@mitb.smu.edu.sg

yanru.xu.2016@mitb.smu.edu.sg

yuchen.wang.2016@mitb.smu.edu.sg