Page 75 - The Economist

The Economist December 9th 2017 Finance and economics 75
2 prone to “overfit”, ie, to finding peculiar Contraceptives and girls’ education
patterns in the specific data they are
trained on that do not hold up in the wider School learning
world. This is especially true of financial
data, he says, because oftheircomparative
paucity. Share-price time series going back In poorcountries, contraception improvesdaughters’ lives as well as mothers’
decades still contain far less information
than, say, the image data used to train Face- EWtasks in developingcountries are Startingin the1960s, these programmes
book’s facial-recognition algorithms. Fas tricky—oras important—as convinc- were introduced in some areas a few
The trick, then, is to take a more ingparents to keep theirdaughters in years earlierthan in others. So research-
thoughtful approach to deployingAI. Tech- school longer. One way ofdoingso is to ers could compare what happened to
nical prowess obviously matters; Sentient make contraceptives available, concludes girls in areas where contraceptives be-
employs a couple of dozen AI experts and a new workingpaperby Kimberly Singer came available when they were very
constantly researches new methods. But Babiarz at Stanford University and four youngwith girls from the same cohorts
business models matter enormously, too. otherresearchers. in areas with no contraceptives.
Sentient started out as a tiny fund a decade Conducted in Malaysia, the study The girls in places with contraceptives
ago, managing only its own founders’ used a happy coincidence ofsurveys stayed in school sixmonths longer, or
money. In the past three years it has ex- goingbackdecades and family-planning about a yearlongerifthey were born
panded into otherapplicationsforAI, such programmes rolled out in a way that afterthe programmes began. Similar
as online shopping and website optimisa- made it possible to measure theireffect. effects have been seen in developing
tion. Only earlier this year did it launch a countries that have specifically aimed to
hedge fund open to outside money, on increase school attendance. But no big
which it hopes to apply the insights changes in school policies accompanied
gleaned elsewhere in its investment arm. the family-planningprogrammes. Nor
Another San Francisco hedge fund that was the extra schoolingbecause these
draws on an even wider pool of expertise, girls had feweryoungersiblings. So the
by virtue of its unusual business model, is boost in school attendance seems linked
Numerai, a firm founded in 2015 that to the availability ofcontraception—for
launched its first fund this autumn. It starts some reason it may have made parents
by taking financial data and then encrypts see the benefits ofeducation.
them so that they are unrecognisable. Its They are considerable. When these
chiefoperatingofficer, MatthewBoyd, says girls reached adulthood, theirjobs were
this turns them into a “pure math pro- much more likely to be paid than those of
blem”. The idea is that this avoids biases women who grew up in otherareas.
creeping into models—and appeals to Val- They were also more likely to have taken
ley types better than the grubby business theirelderly parents in, but not their
ofpickingsecurities. in-laws—a sign that they probably had a
It then runs two-stage competitions for greatersay overfamily decisions.
machine-learningalgorithms that perform All in all, the study suggests that the
best on the data. Some 1,200 data scientists benefits ofcontraceptives in poorcoun-
now take part weekly, competing for virtu- tries may be largerthan thought. A lesson
al prizes (in the fund’s own cryptocur- forpolicy wonks is that it pays to cast a
rency) in the first round and cash prizes in wide net forside-effects when trying to
the second. That structure seeks to encour- Happier families workout whethersomethingworks.
age algorithms that do well at picking win-
ners overtime. The firm takes the results of
the best algorithms, decrypts these results Marijuana and banking in California Although 29 American states allow
back into financial data, and uses the in- sales of marijuana for medical use (or
sights to decide which shares to trade. The From cash to ash medical and recreational use), federal law
fund owes at least as much to crowdsourc- still classifies it as a “schedule 1” drug like
ing, then, as it does to harnessingAI. heroin. Firms handling marijuana pro-
One hedge fund thatdoestoutits mach- ceeds can be prosecuted for money-
ine-dependent model, despite naming it- laundering. Ned Fussell of CannaCraft, a
self after the human brain, is Cerebellum maker of marijuana products, says that a
Capital. Founded as an arbitrage fund in The potbusiness has an unusual fewfirmsopen a bankaccountunderan al-
2008, it started work on a fully AI-run financial problem ternative identity. Butbanksalmostalways
American equity fund in 2016, and ANY marijuana growers in northern find out. So cannabis businesses operate
launched itin April thisyear. The fund uses MCalifornia, America’s biggest source almost exclusively in cash. Many pot farm-
machine learning not just to crunch data of the stuff, had expected this autumn’s ers fled the fires without theirbanknotes.
and come up with strategies. The classifica- harvest to be the largest ever. After all, rec- Chiah Rodriques of Mendocino Gener-
tion system that gauges the relative merits reational marijuana becomes legal in the ations, a cannabis-genetics consultancy,
of these strategies is itself run by machine state in January. Instead, wildfires in Octo- knowstwo dozen people who losta hoard.
learning. But humans do the actual trad- ber—spreading so fast they killed 43 peo- One burnt cabinet had held $250,000.
ing, followingthe algorithm’s instructions. ple—burned up half the marijuana grow- Cheryl Dumont, from a Mendocino Coun-
Howevertheyperform in the longterm, ing in the area’s tri-county “Emerald ty cannabis co-operative, says that of
therefore, one feature of these new AI Triangle” alone and new fires now raging about 20 stashes buried by members or
funds is already clear. At least in investing, will claim more. Some reckon the fires seta neighbours, only one was deep enough to
more artificial intelligence does not neces- record notjustforburntpot, butalso forthe survive. The gold and silver she had in-
sarily mean less ofthe human kind. 7 value ofbanknotes turned to ash. terred melted into a dirt-infused blob. 1