Forecast Performance

Back to What's New
Jump to: [Update #3] [Update #2] [Surf height improvements] [Update #1]
Performance update: Hurricane Erin
Kurt Korte
·
Posted September 16th
As part of our ongoing commitment to improving our forecasts, we are sharing how our LOTUS swell model and our Smart Cams performed—for better or for worse—during the Hurricane Erin swell last month.
Rockaways on August 22nd, 2025.
Forecasting tropical cyclones (e.g., hurricanes) is challenging. Even with major forecast model advancements over the past few decades, significant errors still tend to arise when tracking a storm and its intensity beyond a few days. Erin was a prime example of this persistent issue, with intensity and track errors by government models feeding into LOTUS and contributing to inaccurate predictions.
We’ve chosen to highlight two breaks:
At both of these spots, we compared our Smart Cams’ observations and LOTUS model predictions alongside those of our surf forecasters, who have more than 70 years of combined experience observing and predicting the surf for the US East Coast.
We also analyzed your forecast feedback provided during Erin to assess forecast performance at all spots affected on the US East Coast.

New Smyrna Beach

Central Florida
LOTUS drastically overcalled the size of the easterly swell for August 19 and 20 at New Smyrna Beach, a pattern we saw at many East Coast breaks during Erin.
We made note of these modeling issues in updates that the forecast team published regarding the front end of the swell. We’ve already started work to improve LOTUS for the next East Coast system we get—hopefully soon.
New Smyrna Beach during Hurricane Erin - comparative surf height observations and predictions between Smart Cam observations (yellow), forecaster observations (green), and our forecast (blue).
In contrast, LOTUS under called the NE/ENE swell on August 22–24. This was also well identified ahead of time by our forecast team and communicated in our starred days to watch and Incoming updates as the slower movement and expanding size of Erin signaled to us that the reinforcing swell would be bigger (and better) than modeled.
New Smyrna Beach on August 22nd, 2025.
View New Smyrna Beach Cam

Rockaways

Long Island, New York
Like NSB, LOTUS overcalled the swell building up to the peak, and under called the back end of the swell.
Rockaways during Hurricane Erin - comparative surf height observations and predictions between Smart Cam observations (yellow), forecaster observations (green), and our forecast (blue).
Overall, Rockaways Smart Cam observations and forecaster observations were well-aligned during Erin.
The Smart Cam observed the surf as bigger on Saturday, versus what we identified as the peak of the swell on Friday, but it was still close. It was pumping those days, which is why we dropped a rare Epic rating.
View Rockaways Cam

Your feedback

As part of our review, we also looked at your feedback. We consider your feedback to be one of the best ways to understand how we performed during swells like Erin.
With surf forecasting, perception is reality—if you perceive the forecast as inaccurate, then we have not done a good enough job either predicting the surf, or setting expectations.
Forecast feedback for the Caribbean, US East Coast, and Europe in the days around Hurricane Erin. Collected by Forecast Data Lead, Rob Mitstifer.
Looking at the feedback you sent in, we see a noticeable drop in our accuracy (compared to our baseline) that coincides with the erroneous LOTUS model spikes for New Smyrna Beach and Rockaways detailed above. This reaffirms our need to make improvements to our spot forecasts for events like Erin—and that work has already commenced.
Back to top
Performance update #2
Ben Freeston
·
Posted May 27th
Jump to: [Margaret River] [New Smyrna]
As part of our ongoing commitment to our forecasts, we are surfacing how our LOTUS swell model performed—for better or for worse—for notable swells from last month.
The spots we’ve picked this month couldn’t be more different from each other:
In both cases we use a mix of our forecast model, our Smart Cam observation system and our team of trained surf forecasters to predict, check, validate and improve.

How to read the data

Below we highlight three key locations and events from recent swell activity to give a flavor of our accuracy and the issues we see with our forecasts.
We validate by comparing camera and forecaster derived measurements of individual wave height averaged over 20 minutes to our hourly forecast. You'll see this data in the graphs below:

Margaret River

Western Australia
With our role as official forecaster for the WSL tour stop at Margaret River, it’s interesting to check in on forecast performance here over the last month.
Recent error distribution between forecasted wave heights, Smart Cam observed wave heights, and observed surf heights from our forecasters at Margaret River.
We have a Smart Cam measuring waves, and with consistent surf, the spot is a great test for its huge range of potential size and conditions.
May at Margaret River – comparative surf height observations between actual surf heights (green dots) and our forecast (blue line).
You’ll see from the graph a generally great fit between actual surf heights (green dots) and our forecast (blue line). There’s a slight tendency to overcall here although there are a couple of factors affecting this:
View Margaret River Cam

New Smyrna

Florida, United States
Florida's swell magnet is an interesting summer test bed for our different forecast systems.
Recent error distribution between forecasted wave heights, Smart Cam observed wave heights, and observed surf heights from our forecasters at New Smyrna.
Here, error is low throughout the month, whether we compare Smart Cam to forecast or to our forecast team observations. Average errors are less than half a foot and never more than a foot.
May at New Smyrna – comparative surf height observations between Smart Cam (green dots), forecaster observations (pink dots), and our forecast (blue line).
There are a couple of little moments though where we see divergence:
View New Smyrna Cam
Back to top
Surf height model forecast improvements
Michael Reinwald
·
Posted May 27th
As part of our ongoing effort to improve our spot-specific forecasts, our team recently updated our LOTUS model at over 100 spots around the world.
These updates will improve the accuracy of our forecasted surf heights at each of these spots.

How do we improve our forecast model?

Our forecasters have made (and continue to make) condition and surf height observations at spots around the world, every day for the last 40 years.
These observations made by our team are the gold standard for figuring out what's really going on at your surf spots.
We regularly compare this treasure trove of daily forecaster observations to what our LOTUS forecast model predicted, allowing us to identify opportunities to improve how our model is forecasting breaking wave heights at each spot.

Spot-level improvements

Avalon 10-14th, NJ, United States
Let's take a look at the process of one of our recent updates to improve breaking wave height predictions at Avalon 10-14th in New Jersey:

First, we assess statistical bias and fix systematic errors to make adjustments to the model.

We compare historic forecaster observations of the surf height to breaking wave height predictions to understand how off our model is, and for which swells we are under or over-calling.
LOTUS breaking wave height predictions before (purple dashed line) and after (black solid line) adjustments during Fall 2024 at Avalon 10-14th.
We underperformed for some of the swells during this period, tending to under-call surf heights compared to observations (red dots).
Our adjustments bring the model outputs closer to ground truth—the observations our forecasters made.

After adjusting the model, we check that our changes reduce misses across all swell constellations and height ranges.

Visualization of significant error margins for specific surf height bands during the Fall 2024 time range at Avalon 10-14th.
The outcome is positive: We were able to reduce error across all surf height bands and reduced the overall mean-average error from 1.56ft to 0.51ft, a 68% improvement.
In other words, with these updates, the average error of predicted surf heights will be less than the height of a surfer's head.

We review the spot-level adjustments with our forecasters before updating to make sure adjustments account for all spot-specific factors.

We only push updates after confirming that the optimizations will work with spot-level nuances, including seasonal fluctuations and swell characteristics, recent changes of the beach’s bathymetry, tides, etc.
We only launch these changes where we have enough significant data and confidence that we are reliably and significantly improving the forecast outputs.

What's next?

As we continue to review and make these updates periodically to spots around the world, we’ll let you know when more adjustments are made.
We’re beginning to incorporate Smart Cam observation data, which will allow us to increase the amount of real-time observations we can make at a spot - from 1-2 observations a day to observations up to every 10 minutes—so we can even better adjust our LOTUS forecast with more data.
We will also begin looking into accounting for more dynamic systems in the model itself, such as tides.
Back to top
Performance update #1
Ben Freeston
·
Posted April 22nd
Jump to: [Kirra] [Pipeline] [Swami's]

About this update

As part of our ongoing commitment to our forecasts, we are surfacing how our LOTUS swell model performed—for better or for worse—for notable swells from last month.
We run ongoing validations for every single location we observe globally, comparing what we see happen in the ocean with either our forecasters or machine learning Smart Cams that measure surf heights and conditions.
No forecast is perfect, but identifying specific types of error allows us to work out how to most effectively improve our models, something we're constantly working on.

How to read the data

Below we highlight three key locations and events from recent swell activity to give a flavor of our accuracy and the issues we see with our forecasts.
We validate by comparing camera and forecaster derived measurements of individual wave height averaged over 20 minutes to our hourly forecast. You'll see this data in the graphs below:

Kirra

Gold Coast, Australia
Overall a good fit over the month of March between camera and forecast.
March 5th at Kirra – some comfortably 10ft+ faces on the inside and a camera observed average of 11.6ft in the surfzone.
Although the peak of the TC Alfred swell was under-called substantially, the trend was clear. But as can be seen, the model performed less well here. We think this is largely because of the amount of energy in novel swell directions. This event will give us data to make corrections there.
The forecast here has also tracked consistently about a foot lower than the camera measurement over that window. We typically adjust for this bias periodically to ensure the best possible representation.
Pipeline Smart Cam, March 2025
Observational data from our Smart Cam at Kirra. March 2025.
View Kirra Cam

Pipeline

O'ahu, HI, United States
Comparative surf height and wave height observations from our model, camera and forecasters at Pipeline in March 2025
A continued run of late season swell was extremely well resolved by the model. Average error is less than a foot even through the larger days.
We’ve put a lot of work into data assimilation and timings for NPAC swell over the last few years and that continues to be evidenced in the validations.
View Pipeline Cam

Swami's

San Diego, CA, United States
Pipeline Smart Cam, March 2025
Observational data from our Smart Cam at Swami's. March 2025.
March was generally a good month for Swami's with an average error <0.8ft. However, we had a couple of foot overcall on March 2nd.
Our model notably overcalled the observed wave height on March 2nd, 2025.
Looking at the camera derived heights this is likely attributed to the onshore flow during the day, waves breaking early and softer than they would have held up by better conditions.
View Swami's Cam