Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

How I made $500k with machine learning and high frequency trading

jspauld.com

Post reply on HN
Print view View original

241–250 of 320 posts

Re: How I made $500k with machine learning and high frequency trading

#241

Post by ishaanc90 »

DennisP wrote: post #70
yajoe wrote: post #43

Earlier quoted context omitted.

This. I love crazy projects and Show HN's until the cows come home, but this one is dangerous that I must repeat the warning to others. I cannot emphasize how important it is to understand that people who trade using price action ( http://en.wikipedia.org/wiki/Price_action_trading ) are just speculating based on where they expect the price to move. It's no different than people who play Texas Hold'em online and specu…

This is a bit off-topic, but it's actually quite feasible to get a real edge in Hold'em, and it's not just about spotting other people's patterns. To start with, there's simple probability: knowing the odds of making you hand vs. the payoff in the pot, or the chance of winning with various starting hands. This is pretty basic but a lot of low-stakes players screw it up. If you get it right, their mistakes are your ga…

I have done HFT, played heads up semi professionally, and for my bachelors thesis wrote a paper on a PLO playing bot. I studied Alberta's research and it is phenomenal. The parallels that emerge between HFT and a pokerbot is essentially that the architectures of both systems are kind of same and the details are kind of orthogonal. The edge in Hold'em is kind of gone. The 1/2 games right now are as tough as the 25/50 games 5 years ago. PLO is still pretty exploitable though. The same is true with HFT. Most strategies in HFT no longer work but there are definitely ones that still give you a lot of edge-you just have to think harder : )(just like three betting preflop and cbetting the flop doesnt work anymroe).

Re: How I made $500k with machine learning and high frequency trading

#242

Post by fawce »

owyn wrote: post #205
grumps wrote: post #181

Earlier quoted context omitted.

If you use their library/SDK/Framework.... how do you actually trade, do you still need a brokerage? Sorry a bit new to this field.

Yep, almost any of them have an API these days. Your software would make the list of trades which is uploaded and executed... Oh, I forgot about another backtesting framework for python. It has a built in IDE as well. http://www.quantopian.com

I work at quantopian. We are working on broker integration, so you will be able to trade algos you develop live.

Our backtesting engine, Zipline, is opensource - https://github.com/quantopian/zipline

Zipline was unveiled at PyData NYC, and the presentation materials are here: https://app.quantopian.com/posts/hello-from-pydata

Re: How I made $500k with machine learning and high frequency trading

#243

Post by zdwalter »

Sharma wrote: post #61

Trust me, you earned that much because of your luck. Otherwise Andrew Ng would have partnered with another finance professor and they would have been the richest people on earth!! Imagine trading with their expert systems on global markets. I traded stocks and Forex for years and my experience says, it is not for everyone. What ever indicators,discipline or model you follow it is going to work only if you have the ri…

I like to trade Forex using mql4, any suggestion? Thanks.

Re: How I made $500k with machine learning and high frequency trading

#244

Post by beagle3 »

bubbleRefuge wrote: post #228
beagle3 wrote: post #215

Earlier quoted context omitted.

> No. Every business has a risk element, but what makes this gambling is that there is no good or service being produced. I was not aware that this is what defines gambling. And "no service produced" is certainly wrong by accepted economic theory - arbitrageurs provide a price discovery service for everyone; they get rewarded for exposing the inefficient prices, even though it is done through market mechanics rather…

Futures, which I assume was the original poster's instrument of choice, are a zero sum game by definition as every contract is an agreement between two parties: buyer and seller.

Only if you assume all players only ever use futures. But make an interest synthetic contract (short future long underlying) and you're out of the zero sum regime again. And it's enough that one actor is not inside the zero sum regime to make that apply to the whole game.

Again, it's a great approximation most of time and over most time periods and asset classes, but it is NOT axiomatic in the way most people believe it is.

Remember: as long as there is a way to inject or withdraw more capital into the system (through whatever asset class, as they are all interconnected), the sum is not identically zero.

Just assume one of the stocks is a gold mining company that works efficiently. The share value rises, and the shares are redeemable for the gold, without anyone having to lose anything (except mother earth)

Re: How I made $500k with machine learning and high frequency trading

#245

Post by stmartin »

quant123 wrote: post #95

There is an air of either incredibility or sheer jealousy in these comments. Nevertheless, I just wanted to tell the OP that he did a great job. Thanks for sharing. I work in the finance industry as a quantitative software developer, and it certainly is not an easy job for one person to do. In fact, I tried (independent of my professional work) doing this myself, and I ended up losing a lot of money. If people are tr…

Did you mean incredulity instead?

I'm not jealous of the money he made - I'm a sw developer, I have a good salary, my wife's an accountant, and we do well - what I am 'jealous' of is not having the time to do something like this, regardless of the outcome. Clearly, I'm prepared to take the risk... but if you have 2 children, you may think differently about spending your time doing this instead of on your family.

Money cannot buy happiness. It can pave the way to it, but happiness only comes through surrender to another person or a cause greater than what one can accomplish by him or herself, and in no other way. And when it comes, it's an unintended side-effect, rather than something that one can pursue (which is one of the reasons I hate the 1st amendment and the way it's worded).

Re: How I made $500k with machine learning and high frequency trading

#246

Post by confluence »

A lot of people are stating that this is like gambling - it is - but not in the sense that you think. Firstly he doesn't use his entire bankroll on each trade, secondly he goes long-short consistently over very short periods of time, thirdly he's too tiny to actually move markets, and fourthly he is in and out within a day - where his max var. is 100% on thousands of tiny trades. Think of it like this - he has $100 - he bets $10 of that each day. He can blow that $10 - no problems.

Worst case he runs out of capital over a period of weeks.

He can't blow up in the way that you think - but he can have large drawdowns over a period of weeks.

Markets are eventually consistent scalable systems - and that is why we prefer them over central planning. In the medium term they prices things correctly, cheaply and efficiently (decade+).

In the short term however (sub-decade) - they can't price jack.

Markets are inefficient period - if they weren't, well then P=NP and you could just put your NP-hard problems into a market and get back cheap, quick, accurate results. Oh - wait - protein folding is actually harder than that.

There are 2 major ways to make money in the markets. Value-Growth and statistical arb (often high frequency). The former (Buffett) is highly concentrated bets on the future of business (I'm value - long TSLA/GOOG/Samsung). The latter is looking for thousands of small diversified statistically significant correlations above 50% (random guesses) and trading costs between securities/price movements over short time intervals (aka statistical ghosts in the data - RenTech/Shaw).

Both work. Both work well. And will continue to do so as long as markets exist.

Re: How I made $500k with machine learning and high frequency trading

#247

Post by rekkal »

jspaulding wrote: post #52
147 wrote: post #42

Earlier quoted context omitted.

Why couldn't you do it now?

Well I could try. But it's not going to be any easier now than it was in 2010. For four months I tried everything I could think of to keep it profitable but in the end nothing worked so I had to shut it off.

I have two theories why it stopped working. I think the market sped up. Latencies are always getting lower and your strategy that worked at 10 ms didn't work with players that are at 1 ms. Also, you might have been gamed because your strategy was easily predictable or you were putting too many trades through the same broker/exchange/etc.

Re: How I made $500k with machine learning and high frequency trading

#248

Post by kolbe »

jspaulding wrote: post #209
kolbe wrote: post #203

Earlier quoted context omitted.

FWIW, I couldn't see your pnl chart. I develop algorithmic strategies for a living, and my first reaction to reading your post was skepticism. I'm skeptical for two reasons. (1) because your methods are so unconventional in an industry where convention rules, and (2) because of the time frame of your success, which happened to be one of the more impressive market recoveries in history. I can't tell you how many peopl…

Thank you. You are right - I should clarify things by saying my program had no directional bias. It was a 50/50 split of longs/shorts. Are there other stats I'm missing?

Given a max loss of 2k, we already know the Sharpe Ratio was pretty good.

2009-10 was more than just a huge rally, it was also a period where vol and skew were massively mispriced. I know this is high frequency, but like I alluded to, you need to make sure that what you're doing isn't replicating the pnl profile of low frequency strategies.

So, how did you perform relative to vol sellers? From the market bottom to the end of 2010, the max daily loss for a vol seller was about 3x average daily pnl, and >80% winners. So your returns do sound better, but not incredibly.

But, even if you failed to perform as well as vol selling did over the same period, that doesn't negate the strategy's validity. If returns were not correlated, then it's safe to say that you weren't just inadvertently shorting vol.

So, start there, work out a regression comparing your daily returns to someone selling vol. Do the same with moving average strategies. Mocking up a simple market making back test versus an ES beta is hard, but that too would be a something to test against. I don't expect you to do any of this, and I'm not going to bother to either. I'm just saying that a complete discussion of this subject would include that information.

Re: How I made $500k with machine learning and high frequency trading

#250

Post by tatsuke95 »

marshallp wrote: post #31
tarr11 wrote: post #23

This is just glorified gambling. I am not sure what special insight or advantage he had, other than his own model. Every trader has a model. It could have easily been called "how i lost 500k with machine learning". Like gambling, it's easy to manipulate statistics to show that you did well in some period of time. I worked for a large investment bank about 10 years ago, writing trading programs for quant traders who w…

I don't get that. It would be true if he just made a few trades, but the author claimed to be making 2000 trades a day. Over a period of months winning that wouldn't qualify as blind luck.

Definitely not blind luck; but boosted by the fact that SPY was on a massive bull run at that time.

This is really cool, any way you cut it.

Post reply on HN