Constructing real time estimates of Australian consumer
spending using bank transactions

by Matthew Elias

19.09.2022

clay-banks-c2a0TydMlAs-unsplash (1)_edited.jpg

Summary

Economists need timely and accurate measures of consumer spending to be able to monitor the economy and produce policy-relevant research. Consequently, economists are increasingly using bank transactions data to track consumer spending in more timely and cost-effective ways than traditional survey-based measures. This is important particularly during uncertain and large-scale events such as recessions.

But there is an open question as to whether bank transactions data provide a reliable indicator of consumer spending in the population. The e61 Institute maintains a weekly consumer spending indicator, the Spendtracker, using bank transactions provided by the credit bureau, illion. This paper describes the construction of the tracker and assesses its performance against other publicly available indicators of Australian consumer spending.

What do we do?

Using bank transactions provided by credit bureau illion, we construct estimates of Australian consumer spending across a variety of spending categories. The paper outlines how we do this, including necessary considerations for data cleaning, filtering, and aggregating individual transactions. We then compare our results against three existing data sources on consumer spending published by the Australian Bureau of Statistics and the Reserve Bank of Australia. See Figure 1 for an example, comparing Spendtracker estimates against the ABS Retail Trade estimates at the national level.

Figure 1: Spendtracker estimates vs ABS Retail Trade at national level, month on month percentage changes

 

 

 

 

What are our findings?

The e61-illion Spendtracker is highly correlated with the ABS Retail Trade estimates when measured in monthly growth rates at the national level. This may seem surprising as the individuals included in the dataset are not necessarily representative of the Australian population (because they have experienced a `credit event'). The correlation between the Spendtracker and ABS Retail Trade further increases after accounting for demographic skews present in the illion data through weighting on observed characteristics.

Figure 2: Weighting the data increases the performance of our estimates against ABS Retail Trade

 

 

 

The correlation between the Spendtracker and ABS Retail Trade is also present at more granular geographies, such as the state level. State level correlations between the Spendtracker and ABS Retail Trade for monthly growth rates are very similar to the national estimates for most regions.

The Spendtracker appears to produce reliable estimates of consumer spending in close to real-time. Due to the nature of the bank transactions dataset, the sample size builds over time. The final sample size is around 150-200,000 people in each month which is larger than the ~10,000 people that are observed in the data one week after the end of the month. Despite this, the correlation and prediction error for consumer spending growth does not change significantly with sample size.

                                 

 

 

What does it mean for policymakers?

The results suggest that bank transaction data can produce reliable estimates of Australian consumer spending growth, with appropriate data cleaning and weighting.

Furthermore, the estimates can be produced in near real-time, making the Spendtracker more timely than most traditional indicators. This is helpful to policymakers who need timely and accurate information on the state of the economy, particularly during large-scale and fast changing events such as the COVID-19 pandemic. e61 and illion will continue to work together to improve and update the Spendtracker, including new spending categories and experimenting with alternative weighting schemes. In the future we plan to also include estimates at a monthly frequency and on a seasonally adjusted basis.

Want to learn more? 

retail_trade_weighted_MoM_percentage_change_results_edited.jpg
retail_trade_weighted_MoM_percentage_change_results_edited.jpg
retail_trade_weighted_MoM_percentage_change_results_edited.jpg
retail_trade_unweighted_vs_weighted_summary_edited.jpg
retail_trade_lagged_results_bar_graph_edited.jpg
sample_size_after_x_weeks.png

Figure 3A: The sample size changes through time

Figure 3B: Real-time estimates do not possess significantly higher prediction error (Month on month percentage changes)