Does Feature Engineering Improve Meteorological Predictions Across Asia?

The air above us is never still. It shifts, swirls, and reacts in a chaotic dance of heat, moisture, and pressure that determines whether a farmer's crop survives or a city braces for a storm. For decades, scientists have tried to predict these movements using massive computer models that simulate the physics of the atmosphere. While powerful, these simulations are incredibly expensive to run and sometimes struggle with the sheer complexity of local weather. In recent years, a different approach has gained ground: using machine learning. These are computer programs that learn patterns from vast amounts of historical weather data to make future predictions. However, there is a catch. If you feed a machine learning program raw numbers—like the temperature or wind speed recorded at a specific moment—it often misses the bigger picture. It fails to see that today's weather is deeply connected to yesterday's, or that the seasons follow a predictable cycle. To fix this, scientists use a technique called feature engineering. This is the process of manually creating new clues for the computer to study. Instead of just giving it the temperature, they might also give it the temperature from three days ago, or a calculation of how the temperature and humidity interact. The hope is that these extra clues will help the computer understand the weather better. But a lingering question remains: does adding more clues always help, or can it sometimes confuse the machine?

A team of researchers from universities in Iran and across Asia set out to answer this question on a massive scale. They did not look at a single city or a single type of weather; instead, they examined the entire Asian continent, a region as diverse in climate as it is in geography, stretching from the frozen north to the steamy tropics. They gathered daily weather records from 1,634 different weather stations, covering a period of twenty-four years. Their goal was to test fifteen different ways of preparing data for a machine learning model. Some methods were simple, just adding the date or the day of the week. Others were complex, creating new variables by multiplying different weather factors together or calculating how the weather had changed over the last week. They tested these methods against eight different weather variables: temperature, humidity, air pressure, wind speed, cloud cover, rainfall, and the daily high and low temperatures. They used a standard machine learning model, known as a random forest, to see which data preparation method produced the most accurate forecasts.

The results were surprising and revealed that there is no single "best" way to prepare weather data. The researchers found that adding engineered features improved the accuracy of the predictions in less than half of the cases they tested. In fact, for about 42 percent of the station-and-variable combinations, the extra clues made a statistically significant difference. But for the rest, the extra work was either useless or, in some cases, actively harmful. The study showed that the right approach depends entirely on what you are trying to predict and where you are trying to predict it. For example, when forecasting air pressure, the most complex method worked best. By including every possible clue, from the time of year to the pressure readings from the previous few days, the model reduced its prediction errors by 41–42% across the entire continent. This makes sense because air pressure changes slowly and predictably, so the computer benefits from having a long history of data to study.

However, the story was very different for extreme temperatures. When the researchers tried to predict the daily high or low temperatures, adding complex features often made the model perform worse. In nearly half of the cases for minimum temperatures, the simplest approach—using only the raw data without any extra engineering—was the most accurate. The researchers suggest that for these specific extremes, the basic data already contains all the necessary information, and adding more layers of complexity only introduces noise that confuses the computer. Similarly, for humidity and wind speed, simple time-based clues, like knowing the month or the day of the week, were usually enough. The most complex methods offered little to no benefit for these variables.

The location of the weather station also mattered deeply. The researchers discovered a clear pattern moving from the tropics toward the north. In the warm, tropical regions of Southeast Asia and India, the most successful method for predicting temperature and cloud cover involved creating interaction features. These are mathematical combinations that let the computer see how temperature and humidity work together, which is crucial in climates where the land and air interact in complex ways. But as they moved north into colder regions like Siberia and Kazakhstan, these complex interactions became less useful. In these higher latitudes, simpler time-based features were sufficient. The study also found that rainfall was the most difficult variable to improve. No single method worked well everywhere, and the improvements were often small and scattered, reflecting the unpredictable nature of rain itself.

Ultimately, this research provides a clear roadmap for anyone using machine learning to forecast the weather. It proves that the idea of a universal solution is a myth. You cannot simply apply the same set of data tricks to every weather problem and expect it to work. The best strategy is to be specific: use complex, comprehensive data sets for pressure; rely on simple time markers for humidity and wind; and avoid over-complicating things when predicting extreme temperatures. Furthermore, the location of the forecast matters as much as the variable itself. What works in a tropical rainforest may fail in a temperate plain. By understanding these nuances, scientists can build models that are not just smarter, but more honest about what the data can and cannot tell them. The study concludes that the future of weather prediction lies not in adding more complexity for the sake of it, but in carefully choosing the right tools for the right place and the right job.