Deep learning fails to forecast fresh produce price surges

Pear price forecast error reaches 128.7 percentage points; supply shocks go uncaptured

Some improvement seen in processed foods; better farm-level data seen as urgent

Photo is unrelated to the article. [Getty Images Bank]
Photo is unrelated to the article. [Getty Images Bank]

Even with AI, predicting sharp swings in fruit and vegetable prices remains out of reach, a new study has found. While deep learning improved forecast accuracy for some categories such as processed foods, it showed little advantage over conventional methods for fresh produce. Researchers say the core problem lies not in the technology itself but in a lack of data — particularly advance information on crop yields and shipment volumes — that could signal price movements before they occur.

According to a report released Friday by the Korea Rural Economic Institute titled "Deep Learning and AI-Based Analysis and Forecasting of Agricultural Food Price Volatility," researchers selected 566 items from 581 original series in the consumer price index that met their analytical criteria and forecast year-on-year inflation rates for the current month and the following month.

The analysis drew on 11 categories of publicly available monthly data, including producer prices, import and export prices, domestic supply prices, international raw materials and food prices, exchange rates, and weather conditions. Researchers built a deep learning model capable of learning changes over time and relationships among variables, then compared its accuracy against conventional forecasting methods.

Results varied sharply by product category. For the overall consumer price index, the deep learning model's mean absolute error fell from 0.2925 percentage points to 0.2797 percentage points — a 4.37 percent reduction in forecast error compared with conventional methods.

For fresh produce, however, AI showed no clear improvement. The model was particularly prone to lagging behind actual price movements during periods of sharp increases or decreases.

Pears were a striking example. Analyzing the evaluation period from December 2023 to November 2025, researchers found that the forecast made in August 2024 for September pear consumer price inflation diverged from the actual figure by as much as 128.7 percentage points. A supply shortfall caused by poor harvests the previous year drove up prices, but monthly public price statistics alone were not enough to capture that in advance.

Napa cabbage also showed a maximum forecast error of 63.7 percentage points for the following month's inflation rate. Apples had a relatively high correlation coefficient of 0.82 between forecast and actual values, yet the error improvement rate compared with conventional methods was actually negative, at minus 0.82 percent.

A high correlation between forecast and actual values does not mean the model correctly anticipated price spikes or drops, researchers noted. For fresh produce, they concluded, AI failed to capture the direction and magnitude of price changes in a timely manner.

Processed foods told a different story. The forecast error for the following month's flour inflation rate fell 8.60 percent compared with conventional methods, while the figure for restaurant-served ramyun dropped 18.13 percent. Among fresh items, bean sprouts — whose supply is relatively stable — saw a current-month forecast error improvement of 8.42 percent, and garlic improved by 2.88 percent.

Researchers said the key factor determining forecast accuracy was not simply whether an item was fresh or processed, but whether supply information affecting prices could be captured in advance through available statistics.

Overall, deep learning alone did not consistently outperform conventional methods. Across the 1,132 forecast results analyzed, the average improvement rate was minus 0.36 percent — meaning the AI model actually underperformed existing approaches on average.

Combining deep learning forecasts with traditional time-series models in equal measure, however, reduced errors across all 11 food product groups in the evaluation period compared with the conventional baseline. The overall forecast error improvement rate for food was 5.8 percent. Researchers assessed the hybrid approach as a viable way to compensate for the limitations of AI-only forecasting.

"This research shows that agricultural food price forecasting systems need to be tailored to the characteristics of each product category," said Kim Sang-hyo, a research fellow at the institute. "For processed foods, where producer prices and cost changes are reflected in final prices, combining existing statistics with AI can improve forecasting. But for fruits and vegetables, more direct supply information is needed — including crop conditions, weather by growing region, storage volumes and wholesale market arrivals."

Kim also noted that national monthly weather statistics alone are insufficient to capture growing conditions at individual production sites. "To forecast sharp price swings in fresh produce, we need to expand crop and shipment data by growing region and build forecasting systems suited to the characteristics of each product," Kim said.


adastra@heraldcorp.com