Spatio-temporal modelling of provincial Human Development Index in Indonesia: a comparison of six forecasting approaches
This study evaluates six spatio-temporal models for forecasting Indonesia's provincial Human Development Index, finding that simple local and cluster-adjusted approaches outperform complex deep learning architectures in short-term predictions, particularly when accounting for administrative boundary changes and limited training data.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Measuring how well people live in a specific place is often reduced to a single number, a score that tries to capture health, education, and the ability to afford a decent life. In Indonesia, this score is called the Human Development Index, and it is calculated for every province. For government planners, this number is more than just a statistic; it is a compass. It tells them where progress is happening and where it is stalling, guiding decisions on where to build schools, hospitals, or roads. But a compass is only useful if it points in the right direction. If the map itself changes while you are trying to navigate, the compass becomes confusing. This is the central challenge faced by researchers studying Indonesia: the country is developing rapidly, but its administrative borders are also shifting, creating a moving target for anyone trying to predict the future.
The core of the problem lies in how data is collected. When a large province is split into several smaller ones, the numbers for the original province no longer represent the same land area. A sudden jump in a development score might look like a miracle of progress, when in reality, it is just a change in the boundaries of the map. This makes it incredibly difficult to tell if a region is truly improving or if the numbers are just shifting because the definition of the region has changed. To make matters more complex, the data arrives only once a year, offering a very short history to learn from. Researchers have long wondered if the most advanced computer models, capable of finding hidden patterns in massive amounts of data, could solve this puzzle better than simple, straightforward methods.
A team of researchers at IPB University set out to test this question by comparing six different ways to forecast the Human Development Index for the coming year across all thirty-four provinces of Indonesia. They gathered data from 2017 to 2023, a period that included a major administrative change in the eastern part of the country where new provinces were created. The researchers wanted to see if complex artificial intelligence systems could predict next year's scores better than a simple method that just assumes next year will look like this year. They also wanted to understand how the changing borders affected the accuracy of these predictions.
The study revealed a surprising truth about forecasting in a changing landscape. The most sophisticated tools, which use deep learning and complex network structures to find patterns, did not perform better than the simple baseline. In fact, for the years where the borders remained stable, the best results came from two much simpler approaches. One method, which groups provinces with similar economic and social profiles, added a small, calculated adjustment to last year's score. The other method, which looks at how neighboring provinces influence each other, did the same. Both of these simple methods outperformed the complex artificial intelligence models, which struggled to learn from the limited number of years available.
The reason the complex models failed is tied to the nature of the data. These advanced systems are designed to learn from vast amounts of information collected over long periods, such as traffic patterns or weather data recorded every minute. However, the Human Development Index is only measured once a year, giving the models very few opportunities to learn how things change over time. When the researchers tested the models on a stable set of provinces where the borders did not change, the simple methods produced errors of less than 0.33 points on the index scale. The complex models, by contrast, produced errors that were more than double that size, or even worse. The artificial intelligence models were able to remember the general ranking of which provinces were rich and which were poor, but they failed to predict the small, year-to-year changes that matter most for planning.
A critical part of the study involved dealing with the administrative changes in the eastern part of the country. When the researchers included the provinces that had been split, the data showed a massive, unrealistic jump in development scores that no model could have predicted. This was not a failure of the models to learn, but a failure of the data to remain comparable. The researchers found that when they excluded these changing provinces from the final test, the simple models shone even brighter. This highlighted a crucial lesson for anyone using data to plan for the future: if the map changes, the numbers change, and no amount of mathematical complexity can fix a broken comparison. The best way to handle this is to be aware of the boundaries and to use simple, transparent methods that can be easily understood and checked.
The researchers also mapped out where the predictions went wrong. They found that the errors were not scattered randomly across the country but were concentrated in specific areas, particularly in the eastern islands where development is more complex and data is harder to collect. The simple models made small mistakes that were spread out, while the complex models made large, specific errors in places like Jakarta and the eastern provinces. This suggests that for regional planning, it is better to know exactly where a model might fail than to have a high average score that hides those failures. The study concluded that for short-term planning in Indonesia, a modest, well-understood approach is more reliable than a sophisticated black box. The most effective strategy is to start with what is known, make small adjustments based on local patterns, and always keep a close watch on how the map itself is changing.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.