Regime Dependent Machine Learning and the Limits of Algorithmic Trading in the Dhaka Stock Exchange

In the world of finance, there is a persistent belief that the future price of a stock can be predicted if one simply gathers enough data and applies the right mathematical tools. This idea rests on the concept that markets are not entirely random; rather, they move in distinct patterns or "regimes," shifting between periods of calm, high volatility, and stress. If an investor could identify these shifts as they happen, the logic goes, they could adjust their strategy to buy when the market is quiet and sell when it is chaotic, thereby outsmarting the average participant. This approach has become a cornerstone of modern algorithmic trading, where computers use complex machine learning to scan vast amounts of historical data, looking for signals that human traders might miss. The question that drives much of this research is whether these sophisticated systems can actually turn those signals into real, lasting profit, especially in markets that are less developed and more unpredictable than those in the West.

A recent study focused on the Dhaka Stock Exchange in Bangladesh sought to answer this question with rigorous precision. The researcher, Md. Shakil Sikder, did not simply test a single model on a handful of companies. Instead, he constructed three separate groups of stocks to ensure the results were not a fluke. One group consisted of twelve well-known, liquid companies carefully selected by the author. A second group was a random selection of six companies, chosen by a computer to remove any human bias in the selection process. A third, much larger group of twenty-four additional stocks was added later to test if the findings held true across a wider range of industries, eventually covering forty-two securities across seventeen different sectors. The goal was to see if a specific type of machine learning, which conditions its predictions on the current market "regime," could consistently beat a simple, baseline strategy of assuming the market will do nothing more than repeat its recent average.

The study employed five different forecasting models, ranging from basic statistical methods to advanced artificial intelligence systems like neural networks. These models were fed thirteen different technical indicators, such as past price movements, trading volume, and momentum, along with a signal indicating whether the market was in a normal, high-volatility, or stressful state. The researchers used a strict testing method called walk-forward validation, which simulates real-time trading by training the model on past data and then testing it on future data it has never seen, ensuring that no information leaked from the future into the past. They also accounted for the specific costs of trading in Bangladesh, including brokerage fees, taxes, and the slippage that occurs when trying to buy or sell large amounts of stock without moving the price against oneself.

The results were stark and consistent across all three groups of stocks. None of the sophisticated machine learning models managed to outperform the simple benchmark. In fact, the most complex model, a gradient-boosted tree algorithm known as LightGBM, performed significantly worse than the baseline approach on every single panel. When the researchers tried to use the regime information to guide trading decisions, the strategies failed to generate any profit. Instead, the trading strategies lost the vast majority of the capital they started with. The study found that the underlying signal the models were trying to exploit simply did not exist in a way that could be turned into money after costs were taken into account. Even when the researchers tried to reduce the number of trades to avoid fees, the remaining returns were still negative, indicating that the problem was not just the cost of trading, but the lack of a genuine predictive signal.

A particularly revealing finding concerned how the computer models interpreted the unique mechanics of the Dhaka Stock Exchange. The exchange has a regulatory mechanism that sets a floor price, preventing stocks from falling below a certain level. During periods of extreme stress, this mechanism can cause trading to halt or prices to remain flat for extended periods. The study discovered that the machine learning models, which rely on volatility to detect market stress, misinterpreted these flat-price freezes. Because the price did not move, the models calculated zero volatility and labeled these dangerous periods as "low risk" or "normal." This error meant the models failed to protect investors during the very times they were needed most. The researchers tested a rule to exclude these frozen days from trading, but this adjustment only made the performance worse, as the costs of rebalancing the portfolio outweighed any potential benefit.

The study also addressed a common pitfall in financial research: the difference between what a model knows after the fact and what it knows in real time. The researchers found that the computer's ability to identify market regimes after seeing the entire history of data was vastly different from its ability to identify those regimes as they were happening. When looking back, the model could correctly label a day as stressful, but when making a decision in the moment, it was wrong more than half the time. This gap between hindsight and real-time prediction is a critical reason why many backtested strategies fail when applied to live markets. The study confirmed that even with advanced statistical corrections and multiple layers of validation, the strategies did not demonstrate any skill that could not be explained by random chance.

Ultimately, the research suggests that for the Dhaka Stock Exchange, the market behaves in a way that is consistent with the idea of weak-form efficiency. This means that past price movements and trading volumes do not provide a reliable roadmap for future gains, at least not in a way that can be exploited by current machine learning techniques. The study concludes that the assumption that regime-aware machine learning can be easily transferred to frontier markets with unique regulatory features is flawed. The specific interaction between volatility-based detection and the floor-price mechanism creates a blind spot that algorithms cannot easily overcome. The findings serve as a cautionary tale for practitioners and researchers, demonstrating that in markets with thin liquidity and specific regulatory constraints, the most sophisticated tools may still be unable to find an edge where none exists. The evidence points to a market where the only reliable strategy, statistically speaking, remains the passive one of holding stocks over the long term, rather than attempting to time the market with complex algorithms.