A country’s inflation rate can provide a sense of the health of its economy, or in everyday terms — when a country’s money is starting to gain or lose value. Inflation is the average change in percentage based on a general price index, but in many cases an increase or decrease only becomes evident on a monthly and annual basis. This lag delays a government’s ability to take a swift course of action. Pulse Lab Jakarta and the Directorate of State Finance and Monetary Analysis within the Ministry of National Development Planning have been developing a fit-for-purpose inflation monitoring tool, which can generate same-day inflation information to advise decision and policy makers working in the economic policy domain. Here we describe how the nowcasting inflation tool was developed, its features and practical uses.
Why ‘Nowcast’ Inflation?
An increase in a country’s inflation rate indicates rising prices, which means that with the same amount of money people are not able to purchase what they used to. This is why governments and central banks around the world set broad targets and adopt monetary policy frameworks in an attempt to control domestic inflation. For instance, in 2018 the inflation target set by the Government of Indonesia was 3.5±1 per cent. That is to say, an inflation rate below 2.5 per cent would indicate stagnation within the country’s economy, while an inflation rate above 4.5 per cent would signal economic instability, especially with respect to commodity prices.
Nevertheless, one of the challenges with controlling the national inflation rate exists inherently in the process of measuring inflation — the inflation rate is typically measured after the period in question has passed. As an example, the inflation rate for this April will be calculated and released by Indonesia’s central statistics bureau (Statistics Indonesia) during the first week of May. While a few days lag may seem insignificant for consumers, at a national scale a negative outcome could have an adverse effect on the population. Nowcasting therefore enables daily assessments to ensure that the Government is informed in a timely manner and can plan interventions to cope with such an outcome.
Selecting the Data Sets
Initial discussions between our data science team and representatives from the Directorate of State Finance and Monetary Analysis about the possibility of developing a nowcasting model began in 2017. We considered several traditional and non-traditional data sets to investigate whether they could aid near real-time orientation. For instance, we tested whether sentiments shared by Indonesians on social media regarding the price of certain commodities could be used to help infer the inflation rate, as well as whether information from mass media could serve as a proxy.
After extensive exploratory data analysis, the data set from Bank Indonesia’s Strategic Food Commodities Price Information Centre showed some positive results. This data set contains daily-updated price data of ten strategically selected commodities: rice, red chillies, cayenne peppers, shallots, onions, cooking oil, sugar, beef, eggs and chicken. To gain a comprehensive measure of domestic inflation, we also considered housing prices and automotive prices. Accessing these data sets however was a challenge, and so we used cement prices and automotive sales data from the Census and Economic Information (CEIC) Center as proxies for housing prices and automotive prices, respectively. Other data sets integrated in the model include daily-updated global oil price data (CEIC) and currency rates (Bank Indonesia).
Designing and Testing the Nowcasting Model
We engineered a feature set that includes all the data we collected to train our nowcasting model. By transforming daily prices to represent their directional change over the previous 14 days, we trained a gradient boosted trees machine learning model with a linear daily interpolation of actual inflation as the target value. With this approach, we were able to accurately nowcast the rate of inflation despite only having access to daily data for some of the items in the official inflation basket.
The model can successfully ‘nowcast’ the rate of inflation within a running month with 72 per cent accuracy and an average error rate of 4 per cent. The working model was initially tested in the second week of November 2018, and generated a 3.15 per cent year on year inflation rate for November 2018. We compared the model’s result with the official inflation rate published by BPSin the first week of December that showed a rate of 3.12 per cent. Whenever one or two of the data sets were removed, the model showed varying results that were not close to the officially published inflation rates. This suggests that the current model relies heavily on all the data sets being used to prevent any statistical disturbance.
Using Mass Media Information to Enhance the Tool
To enhance the use of the tool, data from a third party that gathers news and media data (including local print and digital media content) from across Indonesia is also integrated. News regarding inflation and related commodities are streamed from this source, and the nowcasting tool then generates word cloud visualisations to provide a snapshot of the context on the ground. While this does not aim to establish a causal relationship, it is useful for generating insights and assessing possible indications, including at a regional level.
For instance, if the inflation rate in West Java Province appears to be high, users can look at word clouds generated for that particular region for possible indicators to conduct further analysis. Seen in the example above, “beras” the Indonesian word for rice appears larger, which may indicate an ongoing issue with rice production in the region that has led to a higher than usual inflation rate for the month. Users can click on a specific word to view and read the list of associated news sources.
The Directorate of State Finance and Monetary Analysis has been organising a series of workshops with universities and researchers throughout Indonesia to gather feedback from the academic community. During recent workshops held at the State University of Medan and the University of North Sumatera, we received valuable inputs from participants, including a suggestion to include the commodities distribution and sales data to complement the pricing data. The team also plans to conduct further testing to see how the model performs in unusual economic scenarios, such as during the aftermath of a natural disaster.
The nowcasting model at this stage looks promising and can be used to produce timely insights on the state of the country’s economy. The model has received recognition from the Bappenas Minister, Bambang Permadi Soemantri Brodjonegoro, and the Minister has encouraged his team to explore additional features that might be relevant for policy makers across the Indonesian Government. The Directorate is also planning to present the inflation model to the National Inflation Control Team and Bank Indonesia as a national planning and strategy tool to help monitor the inflation rate for a running month. This nowcasting tool will be integrated as part of the Bappenas’ Situation Room that will be launched soon.