Every year, scientists essentially place a bet on which flu strains will make people sick six to eight months in the future. The process combines year-round global virus surveillance, laboratory testing, and expert judgment, all compressed into a tight timeline driven by manufacturing deadlines. It’s one of the largest coordinated prediction efforts in public health, and it doesn’t always get it right.
Year-Round Global Surveillance
The foundation of flu vaccine prediction is a massive surveillance network called GISRS, coordinated by the World Health Organization. It includes 138 National Influenza Centres spread across the globe, six WHO Collaborating Centres, and four Essential Regulatory Laboratories. These labs collectively process over a million clinical specimens per year, tracking which flu viruses are circulating, where they’re spreading, and how they’re changing.
Because flu season happens at different times in different hemispheres, there’s always a flu season happening somewhere. When it’s summer in the Northern Hemisphere, scientists are watching what’s circulating in Australia, South America, and other southern regions for clues about what might arrive months later. This staggered timing is a built-in advantage, though the virus doesn’t always cooperate by staying predictable.
How Scientists Analyze the Viruses
Collecting samples is only the first step. The real work happens in the lab, where scientists run two main types of analysis on the viruses they’ve gathered: genetic sequencing and antigenic characterization.
Genetic sequencing maps out the virus’s RNA to track mutations over time. This reveals how quickly a strain is evolving and whether new variants are emerging that might escape existing immunity. But genes alone don’t tell you everything about how the immune system will respond, which is where antigenic characterization comes in.
Antigenic characterization tests how well antibodies (from previous infection or vaccination) recognize a virus. The primary tool is the hemagglutination inhibition (HI) test, which measures whether antibodies can block a flu virus from clumping red blood cells together. If antibodies raised against the current vaccine strain can still neutralize a new circulating virus, the match is good. If they can’t, that’s a signal the vaccine may need updating. The CDC also uses a newer imaging-based test specifically for tracking how H3N2 viruses, which are notoriously fast at evolving, are drifting away from vaccine-induced immunity.
Much of this antibody testing has traditionally relied on ferrets, because their respiratory tracts resemble ours. Scientists infect ferrets with flu, collect the antibodies they produce, and use those antibodies as a reference for comparing viruses. But the ferret model has real limitations. Ferrets used in these studies have never been exposed to flu before, so their immune response is “naive,” while most humans carry a lifetime of accumulated flu exposures that shape how they respond. Research has clearly shown that human antibodies after vaccination react differently to flu viruses than ferret antibodies do, sometimes targeting different parts of the virus entirely. This mismatch has pushed the field toward incorporating more human blood samples into the selection process.
The Decision Meeting
Twice a year, the WHO convenes experts to formally recommend which strains should go into the next season’s vaccine. The Northern Hemisphere meeting takes place in February (the 2026-2027 selection meeting, for example, was held February 23-26, 2026). The Southern Hemisphere meeting typically happens in September. These meetings review all the accumulated surveillance data, lab results, and modeling outputs, then vote on the specific strains to include.
Each flu strain is named using a standardized system: antigenic type, geographic location of isolation, strain number, and year. So A/California/07/2009 (H1N1) refers to an influenza A virus, H1N1 subtype, isolated in California in 2009 as strain number 7. When the WHO announces its vaccine recommendations, it uses these names to specify which viruses the vaccine should target.
For the 2024-2025 season, flu vaccines in the U.S. shifted from four strains to three. The reason: one of the two influenza B lineages, B/Yamagata, hasn’t been detected anywhere in the world since March 2020. An FDA advisory committee unanimously recommended dropping it, and the vaccines now contain two influenza A strains (H1N1 and H3N2) plus one influenza B strain from the Victoria lineage.
Why the Prediction Is Made So Early
The February timing for a flu season that peaks in December or January seems absurdly early, and in many ways it is. But vaccine manufacturing requires six to eight months of lead time. Hundreds of millions of doses need to be produced, tested for safety, packaged, and distributed before the season begins. That production clock forces the strain selection decision months before anyone knows for certain which viruses will dominate.
Most flu vaccines are still grown in chicken eggs, a process that introduces its own problems. When flu viruses replicate in eggs, they can pick up mutations that make them slightly different from the wild virus circulating in humans. These egg-adaptive mutations can reduce how well the vaccine works, and this has been a particular issue with H3N2 strains. In one well-documented case, the egg-based vaccine virus acquired multiple mutations that meaningfully changed its surface proteins, contributing to a poor match that season.
Cell-based vaccines, grown in mammalian cell cultures rather than eggs, avoid this problem and produce viruses more antigenically similar to the ones actually infecting people. But as of the late 2010s, roughly 89 to 90 percent of flu vaccine doses in study populations were still egg-based, with cell-based vaccines making up only about 10 percent. The WHO now issues separate strain recommendations for egg-based and cell-based vaccines, acknowledging that the best candidate virus may differ depending on the manufacturing platform.
How Often the Prediction Is Right
Vaccine effectiveness varies substantially from year to year, and much of that variation comes down to how well the predicted strains match what actually circulates. CDC data shows the range is dramatic. In the 2010-2011 season, when the match was strong, vaccine effectiveness reached 60 percent. In the 2014-2015 season, when a significant mismatch occurred, effectiveness dropped to just 19 percent. Some seasons have been even worse: the 2004-2005 season saw an estimated effectiveness of only 10 percent, with confidence intervals stretching into negative territory, meaning researchers couldn’t even be sure the vaccine provided any benefit at all.
Even a “good” flu vaccine year doesn’t mean 60 percent of vaccinated people are fully protected. It means vaccinated people are roughly 60 percent less likely to need medical care for flu compared to unvaccinated people. And the H3N2 subtype, which causes the most severe disease in older adults, is consistently the hardest to predict and match, partly because it evolves faster than other subtypes and is more prone to egg-adaptation problems.
Computational Tools for Better Predictions
The traditional process relies heavily on expert judgment, and researchers have been developing computational tools to supplement it. One notable example is VaxSeer, an AI system developed at MIT that predicts dominant flu strains and identifies the most protective vaccine candidates months ahead of the WHO’s decision meetings. In a 10-year retrospective evaluation, VaxSeer’s strain recommendations were tested against the WHO’s actual picks for both H3N2 and H1N1 subtypes, and the model’s predictions correlated strongly with real-world vaccine effectiveness data from the CDC, Canada’s sentinel surveillance network, and Europe’s I-MOVE program.
These tools analyze viral evolution patterns computationally, looking at how mutations in key surface proteins are likely to spread through the global population. They can process far more sequence data than human experts can review manually, and they can simulate how different vaccine strain choices might perform against multiple possible evolutionary scenarios. The goal isn’t to replace the expert committees but to give them better inputs and reduce the guesswork in what remains, fundamentally, a prediction about the future.