Sparks fly as Tesla’s $150 billion market rally leaves fund managers racing to catch up

Elon Musk earned more than $1 billion from Tesla’s recent rally. (Reuters)
Short Url
Updated 08 February 2020

Sparks fly as Tesla’s $150 billion market rally leaves fund managers racing to catch up

  • Tesla shares have climbed nearly 320 percent since early June
  • The rally was helped by ramped-up production at Tesla's new car factory in Shanghai

NEW YORK: Pretty much everyone on Wall Street has an opinion about Tesla.

The electric vehicle maker’s stupendous rally in recent months has given shareholders something to cheer about, cost short sellers billions of dollars and vindicated legions of retail investors who have long adored Elon Musk’s company.

Tesla shares have climbed nearly 320 percent since early June, helped by the company's better-than-expected financial results and ramped-up production at its new car factory in Shanghai.

Another factor driving this week’s rally may be fund managers hurrying to raise their allocation of the stock, analysts said.

“A lot of advisors and institutions, they jump in the bandwagon because they don’t want to trail,” said Ross Gerber, president and CEO of Gerber Kawasaki in Santa Monica. “If Tesla goes to $1,000 and they don’t own it, what are they going to tell their clients?”




Tesla China-made Model 3 vehicles are seen during a delivery event at its factory in Shanghai, China on January 7, 2020. (REUTERS/Aly Song/File Photo)

Gerber trimmed his fund’s position in the stock as the company’s valuation soared. He hopes to buy more if the stock falls and said a fair valuation would be about $550.

Retail investors have driven part of the surge, as staunch defenders of Tesla crowd Twitter, Reddit and other web sites.

Among Fidelity Investments customers, Tesla has been the most actively traded stock in recent sessions, with 16,000 buy orders for the electric carmaker's shares. 

The stock is held widely by institutional shareholders. Tesla’s biggest institutional shareholders are Baillie Gifford, Capital World and Vanguard, according to Refinitiv data. It also has an international following. Retail investors in South Korea have been trading Tesla shares at a furious pace in recent weeks, buying and selling $200 million of stock in January, according to the Korea Securities Depository. Volume in November stood at $43 million.

Tesla options positioning is also bullish. According to data from options analytics provider Trade Alert, skew turned deeply negative this week, meaning that demand for calls, used to position for further share gains, has surpassed demand for puts, used to guard against a fall in shares.

That is a departure from the usual dynamic in most stocks, in which options used for downside protection generally command prices higher than those for upside participation.

Tesla’s biggest winner is Musk, who stands to up to $1 billion thanks to Tesla’s recent rally. The company’s market capitalization briefly exceeded $150 billion this week, the second target in his record-breaking compensation package.


Man vs. machine in bid to beat virus

Updated 20 February 2020

Man vs. machine in bid to beat virus

  • Human and artificial intelligence are racing ahead to detect and control outbreaks of infectious disease

BOSTON: Did an artificial-intelligence system beat human doctors in warning the world of a severe coronavirus outbreak in China?

In a narrow sense, yes. But what the humans lacked in sheer speed, they more than made up in finesse.

Early warnings of disease outbreaks can help people and governments to save lives. In the final days of 2019, an AI system in Boston sent out the first global alert about a new viral outbreak in China. But it took human intelligence to recognize the significance of the outbreak and then awaken response from the public health community.

What’s more, the mere mortals produced a similar alert only a half-hour behind the AI systems.

For now, AI-powered disease-alert systems can still resemble car alarms — easily triggered and sometimes ignored. A network of medical experts and sleuths must still do the hard work of sifting through rumors to piece together the fuller picture. It is difficult to say what future AI systems, powered by ever larger datasets on outbreaks, may be able to accomplish.

The first public alert outside China about the novel coronavirus came on Dec. 30 from the automated HealthMap system at Boston Children’s Hospital. At 11:12 p.m. local time, HealthMap sent an alert about unidentified pneumonia cases in the Chinese city of Wuhan. The system, which scans online news and social media reports, ranked the alert’s seriousness as only 3 out of 5. It took days for HealthMap researchers to recognize its importance.

Four hours before the HealthMap notice, New York epidemiologist Marjorie Pollack had already started working on her own public alert, spurred by a growing sense of dread after reading a personal email she received that evening.

“This is being passed around the internet here,” wrote her contact, who linked to a post on the Chinese social media forum Pincong. The post discussed a Wuhan health agency notice and read in part: “Unexplained pneumonia???”

Pollack, deputy editor of the volunteer-led Program for Monitoring Emerging Diseases, known as ProMed, quickly mobilized a team to look into it. ProMed’s more detailed report went out about 30 minutes after the terse HealthMap alert.

Early warning systems that scansocial media, online news articles and government reports for signs of infectious disease outbreaks help inform global agencies such as the World Health Organization — giving international experts a head start when local bureaucratic hurdles and language barriers might otherwise get in the way.

Some systems, including ProMed, rely on human expertise. Others are partly or completely automated.

“These tools can help hold feet to the fire for government agencies,” said John Brownstein, who runs the HealthMap system as chief innovation officer at Boston Children’s Hospital. “It forces people to be more open.”

The last 48 hours of 2019 were a critical time for understanding the new virus and its significance. Earlier on Dec. 30, Wuhan Central Hospital doctor Li Wenliang warned his former classmates about the virus in a social media group — a move that led local authorities to summon him for questioning several hours later.

Li, who died Feb. 7 after contracting the virus, told The New York Times that it would have been better if officials had disclosed information about the epidemic earlier. “There should be more openness and transparency,” he said.

ProMed reports are often incorporated into other outbreak warning systems. including those run by the World Health Organization, the Canadian government and the Toronto startup BlueDot. WHO also pools data from HealthMap and other sources.

Computer systems that scan online reports for information about disease outbreaks rely on natural language processing, the same branch of artificial intelligence that helps answer questions posed to a search engine or digital voice assistant.

But the algorithms can only be as effective as the data they are scouring, said Nita Madhav, CEO of San Francisco-based disease monitoring firm Metabiota, which first
notified its clients about the outbreak in early January.

Madhav said that inconsistency in how different agencies report medical data can stymie algorithms. The text-scanning programs extract keywords from online text, but may fumble when organizations variously report new virus cases, cumulative virus cases, or new cases in a given time interval. The potential for confusion means there is almost always still a person involved in reviewing the data.

“There’s still a bit of human in the loop,” Madhav said.

Andrew Beam, a Harvard University epidemiologist, said that scanning online reports for key words can help reveal trends, but the accuracy depends on the quality of the data. He also notes that these techniques are not so novel.

“There is an art to intelligently scraping web sites,” Beam said. “But it’s also Google’s core technology since the 1990s.”

Google itself started its own Flu Trends service to detect outbreaks in 2008 by looking for patterns in search queries about flu symptoms. Experts criticized it for overestimating flu prevalence. Google shut down the website in 2015 and handed its technology to nonprofit organizations such as HealthMap to use Google data to build their own models.

Google is now working with Brownstein’s team on a similar web-based approach for tracking the geographical spread of the tick-borne Lyme disease.

Scientists are also using big data to model possible routes of early disease transmission.