Structural Break Detection and Regime-Dependent Financial Time-Series Forecasting Using Change-Point Models and Markov Switching: Evidence from Asia-Pacific Equity Markets

Financial markets are not machines that run on a single, unchanging set of rules. Instead, they are living systems that shift between different moods. Sometimes they are calm and predictable; other times, they are chaotic and prone to sudden, violent swings. For decades, economists have tried to build models that can predict these movements, hoping to see the future clearly. But these models face a fundamental problem: they often assume the past is a perfect guide to the future. When a market suddenly changes its nature—perhaps due to a new law, a global crisis, or a shift in investor psychology—a model that clings to old data can become dangerously slow to adapt. It is like trying to navigate a storm using a map drawn for calm seas.

To solve this, researchers have developed two main ways of thinking about market changes. One approach looks for "regime switches," where the market flips between recurring states, like a light switch toggling between calm and turbulent. The other looks for "structural breaks," which are permanent changes in the underlying rules of the game, where the old map is no longer valid at all. A newer, more sophisticated idea combines these: using real-time signals to detect when the rules have changed, and then instantly adjusting the model's memory to focus only on the most recent, relevant history. This paper asks a simple but difficult question: does this smart, adaptive approach actually work better than the established methods when predicting stock prices and risks across the diverse markets of the Asia-Pacific region?

A team of researchers set out to test this idea using data from five major stock markets: India, Japan, Hong Kong, South Korea, and Australia. They gathered nearly two decades of daily price data, covering everything from the global financial crisis to the pandemic and beyond. Their goal was to see if a model that could "hear" a signal of change and immediately forget its old history would outperform a model that kept all its history, even the outdated parts. They built a system that combined a standard market model with a real-time detector. This detector, known as Bayesian online change-point detection, acts like a sensitive alarm clock. It watches the market's volatility and, if it senses a sudden shift in the pattern, it rings an alarm. When the alarm sounds, the model is programmed to shrink its memory, discarding older data that might no longer be relevant, and retraining itself on the fresh information.

The researchers tested this "break-aware" model against a standard version that never shrinks its memory, as well as against several other common forecasting tools. They ran a strict, one-step-ahead experiment over 1,000 days for each market, ensuring that the model could only use information available up to that specific day, just as a real investor would. They also ran computer simulations where they knew the exact answers in advance, to verify that their methods could correctly identify changes when they happened. The simulations showed that their tools were capable of finding the right patterns; the real-world test, however, told a more complicated story.

The results revealed that there is no single "best" way to predict the future of these markets. The evidence of structural breaks varied wildly from country to country. In South Korea, the data suggested three distinct, permanent shifts in how the market behaved. In Japan, the data showed no such permanent breaks at all, even though the market experienced significant turbulence. This finding alone was crucial: it proved that the markets in this region do not share a single, common rhythm. What works for one country cannot be blindly applied to another.

When it came to predicting the actual price of stocks the next day, the smart, adaptive model did not win. In fact, for most markets, the simple models that just looked at the average past performance or the most recent trends performed just as well, or even better. The complex model that tried to constantly reset its memory did not produce more accurate price predictions. In some cases, like the Japanese market, the adaptive model actually performed worse than the simpler alternatives. The researchers found that the extra complexity of constantly forgetting and relearning did not help predict the next day's return.

The story changed slightly when the researchers looked at volatility, or how much prices swing. Here, the traditional models that kept their full history of data generally outperformed the adaptive model. The model that kept shrinking its memory tended to make larger errors in predicting how wild the market would be. It seems that even when the market changes, keeping a long history of data helps the model understand the underlying patterns of risk better than starting over with a blank slate. The "no-reset" model, which never forgot its past, remained highly competitive and, in several markets, significantly superior at forecasting average volatility.

However, the adaptive model found its strongest niche in predicting downside risk, specifically the Value-at-Risk. This is a measure used by banks and investors to estimate the maximum loss they might face on a bad day. In the Indian market and the Australian market, the break-aware model provided a much better calibration of this risk. It correctly predicted the frequency of extreme losses more accurately than the other models. For the Indian market, it predicted the number of bad days almost perfectly. In the Australian market, it also performed very well. But this success was not universal. In Hong Kong, the adaptive model was too conservative, predicting far fewer bad days than actually occurred, while in South Korea, all models struggled to capture the extreme tail of the risk distribution.

The researchers also looked at how severe the losses were on the days when the risk threshold was breached. Here, the adaptive model did not show a clear advantage. In many cases, the traditional models predicted the size of the losses just as well, or better. The study concluded that the value of knowing about a structural break depends entirely on what you are trying to predict and which market you are watching. It is not a magic bullet that improves every forecast.

Ultimately, the paper suggests that the idea of constantly resetting a model's memory based on real-time alarms is a useful tool, but only for specific tasks. It is not a universal upgrade for financial forecasting. The best approach depends on the market: for some, keeping a long history is vital; for others, a shorter, more focused view helps. The study serves as a reminder that financial markets are too diverse and complex for a single, perfect formula. The most effective strategy is to understand the specific nature of the market in question and to choose the tool that fits that specific environment, rather than assuming that a more complex, adaptive system is always the answer.