Quick take: AI-based models are democratizing weather forecasting and research, creating opportunities for investors to assess commodity and macro risks deeper into the future. Though limited by the data they learn from, ML models outperform standard numerical weather models while cutting computational costs tremendously. The most promising applications span weekly to seasonal timescales. When it comes to long-term climate modeling, ML is unlikely to supplant our physics-based toolkit anytime soon.
Some applications of AI make life easier (code assistants; email summaries), some are depressing (the glut of GPT-generated prose across just about every forum; uncanny deepfakes), and a few are genuinely exciting. As an atmospheric scientist I’m biased, but I believe one of the most transformative applications of AI right now is in weather and climate modeling. The innovations of the last few years amount to the greatest revolution in weather modeling in over half a century. Data-driven weather models now outperform traditional numerical models and provide useful predictions deeper into the future. Extended forecast horizons will enable earlier warnings for severe weather events, allowing for life-and-livelihood-saving preparation measures. At the same time, AI has made weather modeling more computationally efficient and accessible for in-house research. In the finance industry, opportunities are arising at commodities desks and beyond.
Credit: Google DeepMind. <https://deepmind.google/blog/gencast-predicts-weather-and-the-risks-of-extreme-conditions-with-sota-accuracy/>
First, how do these models work?
Weather forecasts might seem like a mundane subject to be excited about. For years it has been possible to open your phone and access free forecasts for any city in the world, in built-in apps and across websites. These forecasts come from numerical weather models— complex sets of equations that encode fluid dynamics and thermodynamics. These process-based models are run in ensembles to trace how small perturbations lead the chaotic atmosphere in different directions.
A new generation of weather models is being formulated in a new way: as generative AI systems that learn how weather patterns evolve over time using historical data. Weather forecasting is an ideal candidate for data-based learning, since future weather conditions depend largely on initial conditions (i.e. the weather tomorrow depends on the weather today), and we have an extensive history of realized trajectories to study. In the past two years, these new models have outperformed conventional models for a range of weather phenomena[1][2]. One drawback of data-driven models is that they are opaque compared to physics-based versions, making it hard to diagnose their behavior. Researchers are pursuing “Explainable AI” to probe the simulations. This work aims to build trust in the models and provide insights that could improve our physical understanding of the atmosphere.
Hybrid AI-physics models have also debuted, which unite the advantages of interpretable, deterministic models with the pattern-recognition superpowers of AI. Google’s Neural GCM is a hybrid physics-AI model that can directly ingest observed precipitation data in the training process. That means it avoids the biases of reanalysis data[3], delivering superior forecasts, particularly for rainfall. Remarkably, this model can be run not only for weather prediction, but for multidecadal climate simulations, a feat of “seamless prediction.” For researchers, achieving unified models that perform across timescales would be a major feat, representing a coherent, flexible approach and a culmination of broad research activities that have historically been disconnected.
AI weather models have rapidly improved in recent years. Source: Stephen Rasp. https://raspstephan.github.io/blog/ai-weather-progress/#
Traditional numerical models are bulky and require supercomputers to integrate equations forward in time. That means most people interested in weather forecasts depend on government agencies or specialized vendors. NOAA’s Global Forecasting System (GFS) and the European Center’s ECMWF typically update every 12 hours. Now, you can find an open-source ML model online and kick off forecasts on your laptop. This technological breakthrough is democratizing weather intelligence.
The investment case
When investors think about climate risk we often want to start by understanding how historical weather has interacted with markets. How does severe heat impact company balance sheets? And how has the frequency of summer heat extremes changed over recent decades? Where are things changing fastest, and why? Is this reflected in asset prices? ML weather models can deliver powerful alternative sources of data for exploring questions like these. Models can be run not only for future weather predictions, but to understand past weather events with “hindcasts.”[4] And since these models are computationally light, it is possible to generate very large ensembles that better sample internal variability and low-probability events. The AI-weather models that support ensembles are particularly useful for investors.
In real-time forecasts, large ensembles can improve probabilistic estimates of future temperatures. Will energy demand be higher or lower in Europe? Will US corn yields be above or below expectations? If you can run more simulations, you’ll have a better chance of accurately answering those questions. Researchers can also train models to optimize them for specific purposes. An energy trader might want to predict wind energy production in Germany or anticipate energy demand in Texas, and tolerate a tradeoff in performance elsewhere.
One concern with data-driven models is that since they “learn” from training data, they may be less capable of simulating rare, extreme weather events. This is of particular concern as climate change increases the likelihood of events that are unprecedented within the modern observational era. Impressively, ML models have shown some success in forecasting events that are extremely rare in a particular season or location, but which have occurred elsewhere on Earth[5]. A team at NVIDIA generated a huge hindcast ensemble for the summer of 2023, which allowed them to quantify the intensity of the historic heatwaves seen that summer[6]. They achieved this at four to five orders of magnitude lower computing cost than with traditional models. Understanding and quantifying how extremes are changing across geographies can help investors think beyond shifts in global, annual means, and get a handle on highly impactful tail risks.[7]
Beyond the investment world, this new technology can bolster resilience and adaptation across society. With AI models improving forecasts and lowering compute barriers, researchers are already developing more tailored and usable climate information. Just last year, researchers at the University of Chicago partnered with the Indian Government to disseminate monsoon predictions to farmers using AI forecasts[8]. The forecasts anticipated a mid-season hiatus of monsoon rains last summer, empowering farmers to adjust their crop management decisions.
Forecasts beyond two weeks: ML for subseasonal to seasonal prediction
Anyone who’s planned a ski weekend or a barbeque knows that unfortunately, it’s impossible to know what the weather will be weeks in advance. The atmosphere is chaotic, so it’s hard to get a reliable read more than about ten days ahead. Data-driven approaches using AI are advancing the frontier of prediction to subseasonal-to-seasonal timescales, the notorious ‘predictability desert’ between the traditional realms of weather and climate models. Where weather is driven by initial conditions (i.e. the pressure and temperature fields today), and climate by boundary values (i.e. the solar radiation, atmospheric composition), subseasonal to seasonal conditions are influenced by both and have a large probabilistic component. With seasonal prediction, one is not expecting a detailed picture but a view on how the upcoming months will deviate from the long-term average. An investor might want to know if this fall will be wetter, drier, or typical in Brazil, and then assess what that means for soybean yields and futures prices.
Though operational seasonal forecasts have existed for over twenty years, AI can make a big impact because the baseline is relatively poor, with large model biases on seasonal timescales. AI pattern recognition has shown promise in improving predictability linked to internal variability modes like the Madden Julian Oscillation (MJO)[9]. The Allen Institute’s ACE2 model, trained for weather forecasting, achieves 1-3 month predictions comparable to standard approaches with lower computational demand, despite applying simple persisted (rather than interactive) sea surface temperatures and sea ice anomalies[10].
So how is it possible to improve forecast performance deeper into the future? As numerical models integrate forward, they lose information from the initial conditions as the chaotic system evolves. Using data and machine learning methods to improve skill, rather than deterministic modeling alone, provides a path forward[11][12]. On these longer timescales, land surface feedbacks and ocean processes become sources of predictability. Effective models link the atmosphere to land and ocean conditions and identify complex patterns across geographies in historical data[13]. As discussed earlier, computational gains are also a key factor for improving seasonal prediction. ECMWF’s seasonal forecasts are released monthly, based on a 51-member ensemble[14]. Larger or more frequently updated forecast ensembles can provide an edge, particularly for energy and agricultural commodity positioning.
Limitations and the investment edge
To balance our enthusiasm, we’ll reflect again that these models are new and have not been vetted as rigorously as conventional models. How can they be trusted? To assess model performance, benchmarking exercises like WeatherBench have been established. However, even these evaluations aren’t straightforward to interpret, since AI models can be trained specifically to perform well on these tests, and may not perform robustly across other dimensions. With public agencies like NOAA and ECMWF running ML weather forecasts operationally, we’ll continue to see how these tools stand up to traditional models over time[15][16].
For now, with private companies like Google and NVIDIA developing and sharing open-source models, access to weather modeling tools is broadening. For teams with the relevant expertise, this is a moment to consider the potential applications for investment decisions across asset classes. Extended forecast horizons, large ensemble tail-risk quantification, and commodity-specific model tuning are some of the top use cases for ML weather models. Improving your view of how likely it is that this summer is warmer or colder than usual, or whether the Indian monsoon will be strong or weak, would allow you to anticipate supply and demand drivers in the coming weeks and months. Getting an early glimpse of outlier events creates an opportunity to hedge. Anticipating temperature or precipitation anomalies months in advance would inform decision making in agriculture, water management, the energy sector, and beyond.
As with all research, it’s important to be critical of the models you use and cognizant of their shortcomings. Particularly on climate timescales, where we seek to understand out-of-sample behavior, AI’s transformative potential is less clear. How would a rapid decline in air pollution affect the timing and intensity of monsoon rains in India? How would that impact agricultural output? Those answers can’t be learned from historical data. But for a range of important problems, data-driven models are already delivering. With open-source desktop-ready models, we now have unprecedented access to weather intelligence. While none of these models is perfect, having more tools at our disposal is taking forecasting into the future. Whether you focus on climate macro risk, energy markets, or real assets, data-driven forecasting tools provide new ways to anticipate extreme events and regional risks. And that enables smarter pricing, risk management, and adaptation. Keeping pace with the whirlwind advances in weather modeling is a worthwhile investment.
[1] ‘ECMWF’s AI forecasts become operational’ (25 February 2025) https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational
[2] Liu, C. C., Hsu, K., Peng, M. S., Chen, D. S., Chang, P. L., Hsiao, L. F., ... & Kuo, H. C. (2024). Evaluation of five global AI models for predicting weather in Eastern Asia and Western Pacific. npj Climate and Atmospheric Science, 7(1), 221.
[3] Most AI models train on ERA5 reanalysis data, which assimilates observations with numerical weather models to create continuous gridded data. In this way AI models and numerical models have a circular relationship.
[4] Hindcasts are “backcasts” created by running weather models with historic initial conditions
[5] ‘A.I. Is Quietly Powering a Revolution in Weather Prediction’ Nikola Jones (April 14, 2025) https://e360.yale.edu/features/artificial-intelligence-weather-forecasting
[6] Mahesh, A., D Collins, W., Bonev, B., Brenowitz, N., Cohen, Y., Harrington, P., ... & Willard, J. (2025). Huge ensembles–Part 2: Properties of a huge ensemble of hindcasts generated with spherical Fourier neural operators. Geoscientific Model Development, 18(17), 5605-5633.
[7] Trok, J. T., Barnes, E. A., Davenport, F. V., & Diffenbaugh, N. S. (2024). Machine learning–based extreme event attribution. Science Advances, 10(34), eadl3242.
[8]Farmers in India Are Tracking Monsoon Season With the Help of AI (WSJ) September 15, 2025
[9] Chen, L., Zhong, X., Li, H., Wu, J., Lu, B., Chen, D., ... & Qi, Y. (2024). A machine learning model that outperforms conventional global subseasonal forecast models. Nature Communications, 15(1), 6425.
[10] Kent, C., Scaife, A. A., Dunstone, N. J., Smith, D., Hardiman, S. C., Dunstan, T., & Watt-Meyer, O. (2025). Skilful global seasonal predictions from a machine learning weather model trained on reanalysis data. npj Climate and Atmospheric Science, 8(1), 314.
[11] Kim, H. S., Zhou, S., Bienkowski, A., & Pattipati, K. R. (2025). Subseasonal to Seasonal (S2S) Prediction Algorithms Using Hybrid Machine Learning Techniques. Artificial Intelligence for the Earth Systems, 4(3), e230108.
[12] Landsberg, J. B. (2025). AI-Informed Model Analogs for Subseasonal-to-Seasonal Prediction (Master’s thesis, Colorado State University).
[13] NVIDIA’s Ocean-linked atmosphere model (Ola) targets seasonal prediction with a model that generates realistic tropical waves and ENSO . Wang, C., Pritchard, M. S., Brenowitz, N., Cohen, Y., Bonev, B., Kurth, T., ... & Pathak, J. (2024). Coupled ocean-atmosphere dynamics in a machine learning earth system model. arXiv preprint arXiv:2406.08632.
[14] ‘Seasonal forecasts’ ECMWF https://www.ecmwf.int/en/forecasts/documentation-and-support/seasonal
[15] NOAA recently launched an AI weather forecasting suite, using Google’s Graphcast as the foundation for AIGFS. A 16-day AIGFS forecast runs in 40 minutes and uses 0.3% of the computing power of the operational GFS. https://www.noaa.gov/news-release/noaa-deploys-new-generation-of-ai-driven-global-weather-models.
[16] ‘AIFS: a new ECMWF forecasting system’ https://www.ecmwf.int/en/newsletter/178/news/aifs-new-ecmwf-forecasting-system
