Well, that was fast. Three weeks later and I’m already fuckin’ fired. This was supposed to be my first update from Manhattan but it looks like I’ll be packing up the truck and heading back down south again first thing in the morning. Hell, Marker and Burge wanted me out of corporate housing today! Which is so stupid, because this whole mess wasn’t even my fault. I thought they knew the kind of stuff I’d be tinkering with, but apparently they’re idiots.
Let me take a step back. I met John Marker at the Atlanta Big Data meet-up back in April. In fact, it turns out he’d flown down from NY just to hear my talk on the difficulties of Natural Language Processing and sentiment analysis. I took the group through typical errors made by the Python Natural Language Toolkit in tokenizing and noun phrase chunking. Everyone is training their models on the same corpus of data, so the models you encounter in the wild tend to make the same mistakes. Entity extraction in ambiguous contexts is especially bad. Sure, you’ve got a negative tweet, but is that negative tweet about Jon Stewart or Kristin Stewart?
Midway through my talk I start showing some ways that high frequency trading gets tripped up by these errors. For instance, how one HFT model ended up shorting Berkshire Hathaway stock because of some negative sentiment it detected related to Anne Hathaway getting cast as Selena Kyle in The Dark Knight Rises. The hedge fund running that algorithm got screwed to the tune of twenty million dollars on that error. Sucks to be them.
Well, Marker freaking loved the talk. After I was done, he was glad-handing me, sliding me his business card, offering to take me and my brothers out for a wild night in his town car. The dude was a real skeeze. But you could tell he was loaded. Like private island loaded, and when I looked him up online later that night, sure enough, he didn’t own a private island but he had taken a Russian rocket up to the ISS.
Later that week, I get a call from Marker and Burge. The two guys are on speaker phone in Burge’s office, and they want to pay me fifty grand to come consult on their HFT models just for the summer. Now that’s some good bank! And at this point I’m already rocking a grand of powder a month, so I ain’t working in the grad lab during the hot months in Atlanta. Atlanta should just go ahead and change the city’s name to Ballsweat each July. So this was a perfect opportunity to cool off, and I jumped on it like a bunny on a carrot.
Now, the first night I’m up there Burge shows me the offices on Broadway before taking in a ride in his Vantage over to the fund’s HFT hardware cages at a facility in Jersey. The whole setup was badass. Lots of fiber. Lots of blinking lights. Their racks were raised with lasers under them to detect people who might want to tunnel into the cage from underground! But he was pretty hush hush on what all algorithms they were running. At this point, I was a bit tipsy, so I just straight up pressed Burge for information.
“How can I consult on your algorithms if you won’t let me know what they are?” I asked.
Burge is this big old fat guy with a ring of hair around the side of his head like one of those olive branch crowns Caesar wore. He kind of just pondered the question and scowled a little. He was sweating through his Brooks Brothers Oxford shirt even though the data center was cold as balls. After a while, I thought he just wasn’t going to answer me when finally he nodded to himself as if he’d finished having an internal conversation with god-knows-who.
“Marker may have wanted you to give some input, but I’m just not comfortable yet. So here’s what we’re going to do. I’m going to set you up with your own account. You can use a bit of the fund’s cash. Ten million maybe. And I want you to show us what you can do. Code something we can put in the datacenter here. You do it right, and then we’ll open the kimono.”
Now if there’s one phrase that makes me want to gag it’s “open the kimono,” and the mental image of Burge doing that made me vomit in my mouth a little. But ten million was nothing to sneeze at. It’d be an awesome line in the old resume if nothing else. I played it cool.
“That’s all?” I asked.
Burge looked a little flustered, “You want more? Earn me some money.”
The fat man waved me to follow him back to his car, and he drove me back downtown to the loft where they were putting me up for the summer. Pretty sweet view of the Williamsburg Bridge by the way. When I rolled in the next day sober, I had half forgotten about the conversation until I had M & B’s head of electronic trading ream me before giving me my network credentials.
His name was Milton Arno, and he had some of the nastiest breath I’d ever smelt. Like he was permanently constipated. But it matched his personality, which also seemed constipated. He was really unhappy about some rookie getting to play at the big kid’s table. Also, he was a Cal Tech grad so he thought Georgia Tech students were all second tier. Getting that vibe just lit a fire under me, and as soon as I got funds in my account, I started hacking together code like a Wall Street Jack Kerouac.
The idea was a simple expansion of the talk I’d given a few months prior. I figured if there were bots out there matching stocks to keywords and keywords to sentiment, such as Berkshire Hathaway to “Hathaway” and looking at sentiment across Twitter, blogs, reddit, whatever, then I was going to piggyback on whenever those HFT algorithms got it wrong. I’d buy a stock just before they thought it was going to tank due to negative sentiment. They’d short the stock. The activity of the other algorithms buying shares would momentarily bump up the price. I’d sell my shares slightly higher, and then I’d wait for the other algorithms to get screwed. Honestly, the whole idea was more of a way to mess with other people’s algorithms than anything. Making money myself was just a nice byproduct.
But in order to pull the whole thing off, I needed to sow negative sentiment about a particular topic. It’s hard to sow negative sentiment online when it comes to a Fortune 500 company. People are less likely to believe it. That’s when I noticed that Hilton had scheduled their quarterly earnings call for 10am the next day.
That night I dreamed of Anne Hathaway. The whole thing had come together perfectly.
The next morning I rolled into Marker and Burge at the butt crack of dawn. Hell, I was there before sour-faced Milton. Using some tools I found on Pastebin, I spun up a bunch of Twitter bots and sent them on a follow-to-get-followed spree. I didn’t care about Twitter shutting them down, so long as they could survive the morning.
The whole thing worked like a charm. In six hours, I’d accumulated a total of two hundred thousand followers across my accounts. Simultaneously I scripted them all to tweet at the major gossip rags. The tweets reported with astonishment that Paris Hilton called Hitler “kindof sexy with that mustache.” This wasn’t but 45 minutes before the Hilton earnings call.
Within ten minutes, I’d made it onto some of the minor RSS feeds, and within half an hour I had hit the front page on reddit. From there the news made the big jump to TMZ. Of course, the rumor was dead within four hours, but it didn’t matter.
I’d bought ten million dollars in Hilton shares right before the propaganda barrage. And just like I’d assumed, a bunch of other algorithms had lousy natural language processing, misinterpreted the sentiment for that coming out of the Hilton call, and shorted the stock to a Biblical degree. In fact, the signal had been so strong and the trading so heavy, that my shares rose 5% before I dumped them.
Of course the earnings were fine. Hell, they weren’t half bad for Hilton. The shorting bots had lost big.
And then I got fired.
It happened pretty quick in fact. Turns out, the reason why I had been brought on by Marker in the first place was because dead-breath Milton Arno and his Cal Tech buddies had written their own NLP HFT algorithm. And just like many others, it’d gotten screwed in the Anne Hathaway debacle. But the only fix they’d put on it was to time the language matches with key company events whenever there was doubt. Turns out, an earnings call counts as a key event.
Milton’s algorithm had shorted the stock to the tune of three hundred million. He was fuming. There were shouts and fingers pointed. And then Burge just calmly told me what was what.
“This is an embarrassment. You should have told us what you were going to do before you did Alex,” he said.
“I figured the ten million was mine to play with,” I said. Hell, I still think I’m right on that score.
“You’re out of here,” was all Burge said. Marker looked a bit hesitant and sad. I think he’d wanted to keep me, but he wasn’t going to step on his partner’s toes.
But then something satisfying happened that made it all worthwhile.
Burge turned to Milton.
“You’re out too. Pack up your things.”
The asshole’s mouth just started silently working. I couldn’t help but smile and hum “Fight for California Tech” at the guy on my way out. For me, the whole thing was no big deal. So I lost an internship. Oh well. That’s still job experience, right?
All that to say, I’m headed back to Atlanta in the morning. I hope you all missed me.