Week-long weather forecasts rely on atmospheric dynamics, while seasonal forecasts depend on slow ocean processes. In between lies a window from two weeks to two months where neither source works at full strength. For agriculture, water resource management, and disaster relief planning, this range is critical: a farmer needs to know when to plant, and emergency services need to preposition supplies. Related work in our archive also seeks to extract signal from climate data using neural networks, but often at other scales: some improve solar flare forecasts, others warn of droughts months ahead. This paper stands apart because it systematically tests a suite of methods specifically on the subseasonal horizon and shows which ones actually work.
The authors assembled data on temperature, soil moisture, sea surface temperature, pressure, and geopotential heights from 1986 to 2018 on a 2°×2° grid over the contiguous United States. Spatial fields were compressed using principal component analysis, keeping ten components per variable, and temporal indices were added: El Niño, the North Atlantic Oscillation, and the Madden–Julian oscillation phase. The target was the average temperature for the third and fourth weeks ahead. Models were trained separately for each month of 2017–2018, tuning hyperparameters on the previous five years of the same month. Ten approaches were compared, from linear regression with L1 regularization to tree ensembles and three neural network variants, including an LSTM encoder with a fully connected decoder.
XGBoost gradient boosting achieved the highest spatial cosine similarity, averaging 0.304, while damped persistence, the standard climate baseline, scored 0.201. The Lasso linear model and the LSTM-encoder neural network reached 0.250 and 0.262, respectively. All three methods consistently beat climatology, that is, forecasts based on the 30-year norm. Maps show that coastal and southern regions are predicted better than interior ones, while Montana, Wyoming, and Idaho are difficult for XGBoost but easier for the LSTM model. Feature importance analysis revealed that models most often pick soil moisture and El Niño indices, sometimes sea surface temperature, but almost never atmospheric variables like 500 hPa geopotential height.
This work confirms that at subseasonal timescales, predictability indeed hides in the slow components of the climate system: soil and ocean, not the atmosphere. Unlike neighboring studies that simply apply a single neural network, this one honestly compares a wide range of methods on identical data and finds that simple models with properly prepared features can match deep ones. That's an important guide for practitioners: you don't always need architectural complexity when data are scarce and correlations are strong. It also shows that an ensemble of different models can smooth out their individual weaknesses across regions.
A natural next step is to combine the strengths of different models into an ensemble that works better everywhere. Neural networks, especially encoder–decoder architectures, have room to improve if overfitting can be tamed and historical information used more efficiently. The feature set could also be expanded to include snow cover, vegetation, or ice cover, which may carry additional signal at these lead times.
Better subseasonal forecasts will directly affect agriculture, water management, and disaster relief logistics. The feature selection methodology could also prove useful in other problems with few samples and strong spatiotemporal correlations.
Test whether the discovered patterns hold on other continents and in other periods, and try incorporating climate model output. Also worth studying is how the contribution of different predictability sources varies by season and region.
For climatology, a key open problem is how to extract maximum information from limited observations under strong spatiotemporal correlation. This work shows that even simple models can be useful if features are properly engineered, but it leaves open the question of which physical mechanisms link slow variables to subseasonal temperature anomalies.
The maximal information coefficient between the temperature of the past two weeks and the forecast for the following two weeks was below 0.1 at most locations in the United States. In other words, recent weather says almost nothing about what will happen a month out.