ElysiumLabs
Member
First name (team leader)
Lee
Last name
Marshall
Organisation name
ElysiumLabs
Organisation type
Small & Medium Enterprise or Startup
Organisation location
United States of America
Models
Please note that the model information below reflects the latest details provided by the team and does not show previous versions. Only the model summary questionnaires retain a submission history.
Model name
Argus
Model type
Post-processing
Number of individuals supporting model development:
1-5
Maximum number of Central Processing Units (CPUs) supporting model development or forecast production:
< 8
Maximum number of Graphics Processing Units (GPUs) supporting model development or forecast production:
< 4
How would you best classify the IT system used for model development or forecast production:
Single node system
Model summary questionnaire for model Argus
The list below shows all model summary questionnaires submitted for this model. Participants may submit an updated questionnaire whenever their methodology changes.
Questionnaires are displayed from the most recent to the earliest, with the latest version opened by default.
Which of the following descriptions best represent the overarching design of your forecasting model?
- Post-processing of numerical weather prediction (NWP) data.
- Statistical model focused on generating quintile probabilities.
- Hybrid model that integrates physical simulations with machine learning or statistical techniques.
What techniques did you use to initialise your model? (For example: data sources and processing of initial conditions)
Argus does not run its own weather model, so there is no initialisation step of its own. It starts from the ECMWF extended-range ensemble forecast initialised at 00 UTC each Thursday and uses nothing dated after that time.
If any, what data does your model rely on for real-time forecasting purposes?
The 101-member ECMWF extended-range real-time forecast and its matching reforecasts (11 members, 20 past years), obtained from the ECMWF Data Store, together with the most recent weekly observations and the 20-year quintile climatology provided by the AI Weather Quest. If the ECMWF data arrive late, a calibrated NOAA GEFS forecast is submitted first and replaced once they arrive.
What types of datasets were used for model training? (For example: observational datasets, reanalysis data, NWP outputs or satellite data)
ECMWF extended-range reforecasts (NWP output), and the ERA5-based weekly observations and quintile climatology supplied by the organisers (reanalysis). Real-time forecasts from 2025–2026 were used only to select and check settings. NOAA GEFS reforecasts are used only for the fallback forecast.
Please provide an overview of your final ML/AI model architecture (For example: key design features, specific algorithms or frameworks used, and any pre- or post-processing steps)
Two components are built for each variable and lead time and then blended. The first is a probabilistic bias correction in quintile space: the fraction of ensemble members below each climatological quintile bound is corrected using the average error of that same quantity in reforecasts for nearby calendar dates, with a second term that also uses the two most recent observed weeks. The second is a small convolutional neural network that adjusts the probabilities from a per-grid-point regression, trained on the reforecasts. Temperature and pressure blend both components; precipitation uses the first alone. MJO phase probabilities come from projecting the forecast's tropical winds and outgoing longwave radiation onto the Wheeler–Hendon basis supplied by the organisers. Tropical-storm-day terciles use climatological frequencies. Everything runs on a single desktop computer.
Have you published or presented any work related to this forecasting model? If yes, could you share references or links?
No, I may have something published after tournaments end.
Before submitting your forecasts to the AI Weather Quest, did you validate your model against observational or independent datasets? If so, how?
Yes. Each design choice was selected on one year of real-time forecasts (2025) and then checked against the organisers' observations on three separate later periods, and was kept only if it did not lose on any of them. A set of recent weeks is held back from all decisions and used only to check the final system.
Did you face any challenges during model development, and how did you address them?
Reforecasts have 11 members while the real-time forecast has 101, so statistics learned on one do not always transfer to the other; correcting in probability space rather than fitting regression slopes avoided this. An ECMWF model upgrade in May 2026 changed the model's behaviour, most clearly for precipitation, so precipitation is now calibrated only on reforecasts from the current model version. Matching each reforecast to the correct verifying week across the turn of the year also needed care.
Are there any limitations to your current model that you aim to address in future iterations?
Skill is weakest for temperature in the tropics and polar regions at days 19–25, and for sea-level pressure in the tropics. The system depends on a single dynamical model, and its MJO forecast uses one ensemble member rather than the full ensemble.
Are there any other AI/ML model components or innovations that you wish to highlight?
The whole system runs on one desktop computer without a GPU, and every change is tested with a fixed select-then-confirm protocol before it is used.
Who contributed to the development of this model? Please list all individuals who contributed to this model, along with their specific roles (e.g., data preparation, model architecture, model validation, etc) to acknowledge individual contributions.
Lee Marshall, myself did everything
Model name
ArgusSafe
Model type
Post-processing
Number of individuals supporting model development:
1-5
Maximum number of Central Processing Units (CPUs) supporting model development or forecast production:
< 8
Maximum number of Graphics Processing Units (GPUs) supporting model development or forecast production:
< 4
How would you best classify the IT system used for model development or forecast production:
Single node system
Model summary questionnaire for model ArgusSafe
The list below shows all model summary questionnaires submitted for this model. Participants may submit an updated questionnaire whenever their methodology changes.
Questionnaires are displayed from the most recent to the earliest, with the latest version opened by default.
Which of the following descriptions best represent the overarching design of your forecasting model?
- Post-processing of numerical weather prediction (NWP) data.
- Statistical model focused on generating quintile probabilities.
What techniques did you use to initialise your model? (For example: data sources and processing of initial conditions)
ArgusSafe does not run its own weather model, so there is no initialisation step of its own. It starts from the ECMWF extended-range ensemble forecast initialised at 00 UTC each Thursday and uses nothing dated after that time.
If any, what data does your model rely on for real-time forecasting purposes?
The 101-member ECMWF extended-range real-time forecast and its matching reforecasts (11 members, 20 past years), obtained from the ECMWF Data Store, together with the most recent weekly observations and the 20-year quintile climatology provided by the AI Weather Quest.
What types of datasets were used for model training? (For example: observational datasets, reanalysis data, NWP outputs or satellite data)
ECMWF extended-range reforecasts (NWP output), and the ERA5-based weekly observations and quintile climatology supplied by the organisers (reanalysis). Real-time forecasts from 2025–2026 were used only to select and check setting
Please provide an overview of your final ML/AI model architecture (For example: key design features, specific algorithms or frameworks used, and any pre- or post-processing steps)
ArgusSafe is the purely statistical counterpart to our Argus model. For each variable and lead time it applies a probabilistic bias correction in quintile space: the fraction of ensemble members below each climatological quintile bound is corrected using the average error of that same quantity in reforecasts for nearby calendar dates, with a second term that also uses the two most recent observed weeks. No neural network is used. MJO phase probabilities come from projecting the forecast's tropical winds and outgoing longwave radiation onto the Wheeler–Hendon basis supplied by the organisers. Tropical-storm-day terciles use climatological frequencies. Everything runs on a single desktop computer.
Have you published or presented any work related to this forecasting model? If yes, could you share references or links?
Not at this time, but there may be some work post-competition
Before submitting your forecasts to the AI Weather Quest, did you validate your model against observational or independent datasets? If so, how?
Yes. Each setting was selected on one year of real-time forecasts (2025) and then checked against the organisers' observations on three separate later periods, and was kept only if it did not lose on any of them. A set of recent weeks is held back from all decisions and used only to check the final system.
Did you face any challenges during model development, and how did you address them?
Reforecasts have 11 members while the real-time forecast has 101, so statistics learned on one do not always transfer to the other; correcting in probability space rather than fitting regression slopes avoided this. An ECMWF model upgrade in May 2026 changed the model's behaviour, most clearly for precipitation, so precipitation is calibrated only on reforecasts from the current model version.
Are there any limitations to your current model that you aim to address in future iterations?
Skill is weakest for temperature in the tropics and polar regions at days 19–25, and for sea-level pressure in the tropics. The system depends on a single dynamical model.
Are there any other AI/ML model components or innovations that you wish to highlight?
A simpler, fully statistical version of Argus, run alongside it as a robustness check. It needs no GPU and no model training.
Who contributed to the development of this model? Please list all individuals who contributed to this model, along with their specific roles (e.g., data preparation, model architecture, model validation, etc) to acknowledge individual contributions.
Myself Lee Marshall, did everything