"Brought an Umbrella for Nothing" Another Wrong Weather App?… In Reality, '9 Out of 10 Forecasts Were Correct' [Reading Science]

AI Excels at Calculations, but Struggles with Physical Interpretation of Weather
"Forecasts Are Not Off the Mark?We Just Misread Them"

Recently, despite frequent "autumn rainy seasons," weather apps have failed to fulfill their roles. Rain was forecasted, but the sky remained clear; on days predicted to be sunny, sudden showers poured down. Among commuters who carried umbrellas only to see sunshine all day, complaints like "The Meteorological Administration got it wrong again" are repeated.

At some point, weather apps on smartphones became "unreliable sources of information." In reality, however, it is much more common for us to interpret the language of forecasts differently, rather than the forecasts themselves being "wrong." Meteorologists say, "It's not that the forecasts are inaccurate, but that users interpret the meaning of the forecasts differently."

Naver Weather Alert. Naver Screen Capture

Forecasts May Seem Wrong, but the Data Is Becoming More Accurate

According to statistics from the Korea Meteorological Administration, the average annual accuracy of short-term forecasts last year was 90.0%. The accuracy (ACC) of precipitation prediction during the recent three months with heavy rainfall (June to August 2025) was 88.5%, which is a clear improvement compared to 84.5% in 2024 and 83.8% in 2023. In numbers, this means that forecasts were correct 9 out of 10 times.

Yet, citizens still feel that "forecasts are often wrong." The reason lies in the language of "probabilistic forecasts." A "60% chance of rain" does not mean "it will rain with a 60% probability," but rather, "in similar weather conditions, it rained 6 out of 10 times in the past"-a statistical meaning.

Most users intuitively interpret this as "it will rain." Ultimately, the gap between the mathematical language of forecasts and citizens' intuition makes the perceived error much greater than the actual error.

The Korea Meteorological Administration has recently begun full-scale operation of an artificial intelligence (AI)-based forecasting system. The ultra-short-term precipitation prediction model "NowAlpha" receives precipitation data observed over two hours from ten weather radars nationwide and predicts precipitation intensity for up to six hours ahead at ten-minute intervals. The calculation takes less than 40 seconds, making it more than ten times faster than conventional numerical forecasting models.

For medium-range forecasts, the latest AI models such as "WISDOM," "FourCastNet2," "Pangu-Weather," and "GraphCast" are used to produce predictions in 12-day units at six-hour intervals.

Yoon Seyoung, Professor at the School of Computing (and Kim Jaechul Graduate School of AI) at KAIST, said, "AI forecasting models operate on a data-driven approach that learns patterns based on observational data, rather than directly solving physical equations," adding, "While it takes time to train, once training is complete, new forecasts can be generated in just a few seconds."

However, this approach is structurally difficult to explain "why a forecast was made that way." Professor Yoon pointed out, "Deep learning is a black box structure with more than millions of weights, so it is difficult to clearly identify which input variables influenced the results," noting that "this creates a structural limitation in explaining the causes of forecasts."

"AI Is Fast, but Stability Remains an Issue"... Addressing Uncertainty Is Key

The Korea Meteorological Administration acknowledges this issue. "AI forecasts excel in speed, but in extreme weather conditions such as storms or heavy downpours, their stability can be lacking." The reason is that AI forecasts use a deterministic structure that presents only a "single result."

To address this, the Korea Meteorological Administration is introducing "explainable AI (XAI)" technology, which visualizes the basis of forecasts so they can be logically explained. The agency is also experimenting with an "ensemble" approach, training multiple models simultaneously and synthesizing the results.

In particular, the Korea Meteorological Administration has recently begun developing a medium-range AI model capable of 14-day predictions by applying a "Bayesian neural network-based ensemble structure." This is designed to move beyond the existing deterministic structure and consider both the uncertainty of initial conditions and the variability of data.

A Korea Meteorological Administration official explained, "While AI calculates quickly, it does not always satisfy physical consistency, such as the law of conservation of energy," adding, "We are developing the system to include physical constraints in the training process in the future, so as to ensure both accuracy and stability."

The Reliability of Forecasts Ultimately Depends on 'Data Quality'

The performance of AI forecasts ultimately depends on the quality of input data. According to the Korea Meteorological Administration's analysis, the factors determining forecast accuracy are: quality of observational data (32%), numerical forecasting models (40%), and the competency of forecasters (28%). In other words, one-third of the forecast is determined by "data quality management."

Jo Junghoon, Researcher at the AI Meteorological Research Division of the National Institute of Meteorological Sciences, said, "The accuracy of AI forecasts is directly proportional to the reliability of input data," adding, "If even part of the observational data accumulates errors, the forecast model's bias can increase." He further noted, "Ultimately, the performance of AI is determined more by the consistency and integrity of input data than by the algorithm itself."

Woo Jingyu, Meteorological Forecaster at the Korea Meteorological Administration, also emphasized, "The observational data used in forecasts are combined from various sources-radar, satellite, marine, etc.-along with the background of numerical models," adding, "The process of mutual adjustment and quality control of the entire system is more important than minor errors in individual data sources."

"Forecasts Aren't Wrong-They're Constantly Being Updated"

Weather alerts on smartphones change multiple times a day. This is not evidence that forecasts are wrong, but rather a process of real-time adjustment to improve accuracy.

Forecaster Woo Jingyu explained, "AI forecasting models can be recalculated at short intervals, so whereas forecasts were previously updated twice a day, now they are updated dozens of times," adding, "A changing forecast actually means it is responding sensitively to weather conditions."

A representative from a private weather app also commented, "From the user's perspective, frequent forecast changes may cause anxiety, but it does not mean the forecast has failed-it is a signal that it is being continuously updated," adding, "In the era of AI forecasting, timeliness is becoming a more important evaluation criterion than accuracy."

Weather forecasting is, by nature, a science that deals with uncertainty. AI is increasingly becoming a tool to reduce that uncertainty, but a "never-wrong forecast" still does not exist. However, by uncovering patterns invisible to the human eye, AI enables us to be better prepared for each day. The science of reading the sky is now evolving into the realm of "wisdom in understanding what is wrong."