On his own, a 27-year-old Chinese man built the most accurate COVID-19 prediction model. How good is he?

On his own, a 27-year-old Chinese man built the most accurate COVID-19 prediction model. How good is he?

The 27-year-old Chinese guy became famous in one battle! He developed the most accurate prediction model for the new coronavirus in the United States and overturned professional institutions by himself

At only 27 years old, he was rated as a "new coronavirus data superstar" by Bloomberg.

Why?

The new crown prediction model built by oneself in only one week surpasses in accuracy professional institutions with billions of dollars and decades of experience.

He is Youyang Gu, who has a master's degree in electrical engineering and computer science and a degree in mathematics from MIT.

But it is worth noting that he is a novice in medicine and epidemiology.

His model was even highly praised by Jeremy Howard, the famous data scientist and founder of fast.ai:

The only model that looks reasonable.

He is the only one who actually looks at the data and does it right.

Not only that, his model was also adopted by the US Centers for Disease Control and Prevention.

What kind of prediction model is it?

The time point goes back to the beginning of last year.

At that time, the epidemic had already spread around the world, so the public tried to use modeling to predict the impact of the next epidemic.

Most attention has been directed to the prediction system built by two professional institutions? Imperial College London and the Seattle-based Institute for Health Metrics and Evaluation (IHME).

However, the predictions given by the two institutions are very different:

Imperial College London: By the summer, the number of deaths from the new coronavirus in the United States will reach 2 million .

IHME: The death toll is expected to reach 60,000 by August.

(Later facts proved that the death toll was 160,000.)

Why is there such a huge gap in the forecast data given by the two professional organizations?

This attracted the attention of Youyang Gu, who was only 26 years old at the time.

Although he does not have any experience in medicine or epidemics, he firmly believes that data predictions will come in handy at this time.

So, around mid-April, Youyang Gu spent only a week at home to build his own predictor and a website that could display relevant information.

But the method used by Gu in this process is not so advanced. On the contrary, it is quite simple.

He first considered the relationship between the number of new coronavirus tests, the number of hospitalizations and other factors. However, in the process, Gu discovered that there were inconsistencies in the data provided by various states and the federal government. .

At this point, the question arises? What kind of data is reliable?

Gu believes that the most reliable data seems to be the daily death toll:

Other models use many data sources, but I decided to use the past death toll to predict Future death toll.

As for the reason for doing this, the explanation given by Gu is that "using it as the only input helps filter the signal among the noise."

So, what are the prediction results?

It can be said to be quite accurate.

When the model was first completed, he predicted that 80,000 people would die in the United States on May 9. The actual death toll that day was 79,926.

The forecast data also from IHME is that the number of deaths in 2020 will not exceed 80,000.

Gu also predicted that on May 18, the death toll would reach 90,000; on May 27, the death toll would reach 100,000.

Facts have proved that his two predictions were successful again!

In addition to forecasting accurate numbers, Gu predicted a second wave of large-scale infections and deaths based on the gradual transition from lockdown to openness in many states.

On the day Gu issued such a prediction, what did Trump say? The 60,000 deaths predicted by IHME indicate that the epidemic will soon end

Maybe it is. Because of the accuracy of Gu’s model predictions, more and more people are paying attention to his work.

On Twitter, Gu not only @called various reporters, but also sent emails to epidemiologists to ask them to verify their data.

At the end of April last year, Carl Bergstrom, a famous biologist at the University of Washington, released Gu’s model on Twitter.

Soon after, the U.S. Centers for Disease Control and Prevention also released Gu’s data on its new crown prediction website.

Not only that, as the epidemic developed, Gu, a Chinese immigrant, also participated in regular meetings organized by a team of American experts, everyone wanted to better improve his model.

His website traffic has also shown explosive growth, with millions of people viewing his data every day.

Normally, the data predicted by Gu’s model will be reached within a few weeks and are very close to the actual death toll.

As similar prediction models gradually increase, Nicholas Reich, associate professor in the Department of Biostatistics and Epidemiology at the University of Massachusetts, Amherst, counts 50 such models:

Gu’s models have always been at the forefront.

But in November last year, Gu made a surprising decision to end his prediction mission.

Reich commented on this:

Youyang Gu is a very humble person. When he saw that other people’s models were also doing well, he felt that his work had already Done.

One month before Gu decided to stop the project, he predicted that the number of deaths would reach 231,000 on November 1, while the actual number was 230,995.

But Chris Murray of IHME believes:

The machine learning method used by Gu has a relatively good effect in short-term prediction, but it does not quite understand what is happening in the big picture.

In this regard, Gu did not respond to the evaluation of the model. Instead, he stated this:

I am very grateful to Dr. Chris Murray and his team for their work; without them , I wouldn’t have achieved what I have today.

After taking a break, Gu returned to this career.

This time, his predictions are: How many people in the United States have been infected with the new coronavirus? How quickly will the vaccine be rolled out? When (if possible) will the United States reach herd immunity? ?wait.

His projections suggest that by June of this year, about 61% of the U.S. population should have some form of immunity—either from a vaccine or from a past infection.

Gu has always wanted to find a job where she could have a huge impact on society while avoiding the politics, prejudice, and burdens that sometimes come with large institutions. He believes:

In this field, there are many shortcomings that can be improved by people with my background.

Who is Youyang Gu?

Youyang Gu was born in a Chinese-American immigrant family and grew up in Illinois and California.

Gu has been interested in mathematics and science since he was a child, and it was not until he graduated from high school that he really came into contact with computer science. He was able to enter this industry thanks to his father, who was a computer practitioner.

Gu studied at MIT for his undergraduate and master's degrees, where he received a double bachelor's degree in computer science and mathematics, and a master's degree in computer science.

After graduation, he continued to conduct research in the NLP group of MIT's famous CSAIL laboratory for one year, and published a paper at EMNLP 2016 in the same year.

This was also his first exposure to big data, and he established a statistical model to predict the data.

However, he did not continue academic research, but entered the industry. After leaving MIT, he joined the financial industry and wrote algorithms for high-frequency trading systems.

There, his data modeling skills were further honed, since in financial trading, data must be very quantitative and as accurate as possible.

After that, he entered the sports world and continued to conduct research on big data. This also provides him with a wealth of interdisciplinary experience, allowing him to successfully tackle new areas and understand how to model more accurately.

In his own words, his expertise is using machine learning to understand data, separate signal from noise and make accurate predictions.

When building a model of COVID-19 deaths, he initially considered the relationship between the number of confirmed cases, the number of hospitalizations and other factors. He then discovered that states and the federal government were reporting inconsistent data, with the most reliable numbers being daily deaths.

Gu believes that if the input data quality is very low, then the more data, the worse the output performance will be.

Within a week, he built a simple model based on death data and put the prediction website online.

Since April last year, Gu has volunteered thousands of hours on the project, free of charge.

In an interview with Eric Topol, editor-in-chief of the medical website Medscape, Gu said that he is now working full-time on the new crown prediction website. He has no part-time job and no income. He lives on his past savings.

However, such a public welfare project was criticized by some Twitter netizens, but he still persisted.

Since December, covid19-projections.com has been accepting donations from netizens and has now reached its fundraising goal of US$50,000.

In addition to the number of infections, Gu’s new crown website has a new feature. Since December, covid19-projections.com has been tracking and modeling vaccination rollout and the path to herd immunity.

This month, Gu changed "herd immunity" to "return to normal" because his model predictions indicate that the United States is unlikely to reach theoretical herd immunity in 2021.

What’s the way forward? What are Gu’s career plans after the epidemic is over?

He said it is still too early. Although his current job is to predict the development of the epidemic, it is difficult for him to predict what he will do in three months or one year.

Because of this work, universities and companies around the world have offered him an olive branch.