TEN Project Proposal

Back to Project Group

PROJECT DETAILS

PROJECT POSTER

PROJECT APPLICATION

REPORT

Problem & Motivation

Forecasting time series data requires technical knowledge and specialised software, making it difficult to do for most people. Even for those with the required resources, completing the entire forecasting process requires many complicated steps to be done manually, making it a tedious process. This problem is exacerbated when forecasting for different datasets is required and the entire process is repeated multiple times. Therefore we would like to create a web application for business and government agencies to be able to assist these organisations in predicting the future performance of their operations by simplify the process.

Approach

The application aims to provide users with the following capabilities:

1. To be able to allow users to select their own data for analysis
This will be achieved by dynamically generating the filters based on the columns in the dataset. Taking the Singapore Tourism data as an example, users will be able to filter by Country and Date.

2. To allow users to provide inputs for exploring and understanding their data through the visualisations
During the data discovery and exploration of the time series data, the exploration will be made as realistic as possible. For example, correlogram includes specifying of lags/periods, hence we will give the user the autonomy to explore the data by specifying their own input value.

3. To allow users to generate forecasts quickly and easily
This will be achieved by providing selections and inputs for the users to generate forecasts of different methods.

4. To allow users to compare the performance of the generated forecasting results
This will be achieved by providing various forecasting methods and the results of each forecasting methods such that users can estimate the accuracy of the models.

Related Works

To better understand forecasting models and time series, our team has looked into the forecasting functions in JMP Pro and related applications that were built with R. This is to understand the visulisations that we should use and terms and models used in forecasting.

JMP Pro
Looking at how forecasting is done in JMP:
- Forecasting in JMP includes the following default visualisations: Time series, Autocorrelation and Partial Autocorrelation
- Users can specify lags for autocorrelation
- Using the ARIMA forecasting, the user can specify 3 parameters and generate different combinations
- Remove linear trend with decomposition and differencing

Time Series Forcasting Tutorial App
The app developer has shown various forecasting models and is also developed with ShinyR. For

References

Related Works
1. Exploring Time series and Forecasting: Time Series Forcasting Tutorial App

Data sources
1. CEIC: Data on Visitor arrivals, Revenue and Expenditure
2. Singapore Tourism Board (STB): Data on Hotel statistics and Tourism sector performance

Key Challenges

ShinyR
In order to leverage on the various forecasting models that is available in R and its libraries, our group will be be using ShinyR to develop our web application. ShinyR also provides the platform to create a dashboard.However, our team forsee the learning of R to be a major challenge as none of us has build a web application with ShinyR, and based on intital research, the R syntax is also quite different.

Forecasting Model & Visualisations
As our group is working on providing a platform for forecasting, we needed to understand the various methods that are available and the visualisations that will be useful for the users to identify the hidden patterns in their data.

Milestones

Future Work

1. Comparing multiple forecasting methods & finding the best
Our application will only be showing the results of forecasting for users to evaluate the accuracy of each. However, it will be better if the application could compare and recommend the best, such that users with no prior knowledge of forecasting can easily understand the results.

2. Forecasting of stream data
Our application can work towards forcasting live data source. This would eliminated the need for users to upload their own data manually.

3. Allowing wider range of data formats
Our application will only allow the reading of csv data file. With a wider range of formats, it will be more flexible for the user.

Thought Process

First Tab: Dynamic Filters Filters should be dynamically generated from the data source as user inputs/filters.

First Tab: Time series & Correlogram
To allow users to do the exploration first before moving to the next tab which is applying forecasting models. The visualisations will include time series graph to look at the trend for the past periods and correlogram to understand the correlation between the lags/periods. We want to allow users to explore the data by smoothing the linear trend with decomposition and differencing with different differences. As for the correlogram, it should allow users to specify the number of lags/periods that they would like to see.

Second Tab: Forecasting Models & Results
After the data exploration in the first tab, users are now ready to perform forecasting with the models provided. Similarly, we want to give users the autonomy to select different models and inputs. We would also provide the results for users to evaluate the accurracy of the model.

Comments