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

Jane Street Market Prediction ($100k Kaggle competition)

kaggle.com

Post reply on HN
Print view View original

81–90 of 217 posts

Re: Jane Street Market Prediction ($100k Kaggle competition)

#81

Post by x87678r »

usmannk wrote: post #24

As a frequent Kaggler (perhaps too frequent... it's a bit addicting, in a way I'm sure others on HN will understand), I was fairly intrigued to see this one pop up in the competition list a few days ago. Finance shops have tried their hand at Kaggle before, but I think they've normally been out of their domain. e.g. Two Sigma recently did a reinforcement learning game competition. I'd caution the HN crowd not to expe…

They even say in the instructions:

Admittedly, this challenge far oversimplifies the depth of the quantitative problems Jane Streeters work on daily, and Jane Street is happy with the performance of its existing trading model for this particular question.

Re: Jane Street Market Prediction ($100k Kaggle competition)

#82

Post by homie »

npmisdown wrote: post #79

I'm sorry, if you could build a model to predict markets, why will you post in to Kaggle to get $40k in prize instead of applying this model to your own broker account?

Mostly because it’s impossible to accurately predict the market - and this is just a competition to see who can build the best model.

Re: Jane Street Market Prediction ($100k Kaggle competition)

#84

Post by Spinnaker_ »

usmannk wrote: post #24

As a frequent Kaggler (perhaps too frequent... it's a bit addicting, in a way I'm sure others on HN will understand), I was fairly intrigued to see this one pop up in the competition list a few days ago. Finance shops have tried their hand at Kaggle before, but I think they've normally been out of their domain. e.g. Two Sigma recently did a reinforcement learning game competition. I'd caution the HN crowd not to expe…

We should also remember that Jane Street is primarily an ETF market maker. Their main business isn't betting on prices of stocks or managing a portfolio.

I've only taken a quick look at the data, but the problem doesn't seem to be focused on their core competencies, but instead is much more general.

Re: Jane Street Market Prediction ($100k Kaggle competition)

#85

Post by justjonathan »

minimaxir wrote: post #40
1helloworld1 wrote: post #38

Isn't it pretty well known in the finance world that using stale public information to predict the market is a fool's errand? Unless you have some kind of specialized non-public data (e.g satellite images of number of cars parked outside parking malls, number of cargo ships moving in and out), trying to predict the market with historical data does worse than "Just give me some monkeys, darts and a dart board".

Granted, a typical Kaggle metagame-that-is-technically-against-the-rules is to use data from outside the dataset, which is one of the reasons winners have to be validated.

From: kaggle.com/c/jane-street-market-prediction/overview/code-requirements

"Freely & publicly available external data is allowed, including pre-trained models"

Re: Jane Street Market Prediction ($100k Kaggle competition)

#86

Post by tikhonj »

npmisdown wrote: post #79

I'm sorry, if you could build a model to predict markets, why will you post in to Kaggle to get $40k in prize instead of applying this model to your own broker account?

It is much harder to turn a model into a profitable trading strategy than people realize. Apart from transaction costs, risk management and market impact there are also a lot of small operational details which can make or break your execution. One example I vaguely recall was that the details of how a specific foreign exchange conducted its closing auction could make a substantial difference to a strategy that involved executing there alongside other trading venues.

The payoff for getting these operational details right or wrong is massively asymmetrical. If you get everything right, you'll only do as well as your model lets you. But if you get anything wrong, you run a real chance of losing far more money than you could have hoped to make!

Even just validating your strategy on historical data (ie back-testing) is harder than it sounds. If you make a mistake that leaks information to the code you're testing, you can end up with a much rosier return and risk profile than you really have. Another way to lose money when you go put your model into action.

If you get over these challenges and run your strategy successfully for a while, other market participants are going to start adjusting against it and you have to adjust in turn. You can't just "set and forget".

I should note that I am far from an expert on any of this, though! I just know enough to not trade with serious money—my real savings are all in index funds I don't touch, thank you very much :).

Re: Jane Street Market Prediction ($100k Kaggle competition)

#87

Post by Spinnaker_ »

toomuchtodo wrote: post #4
usmannk wrote: post #3

Earlier quoted context omitted.

You don't have to (and certainly won't) beat all of Jane Street. The goal is to beat everyone else on Kaggle. A still difficult but much more accomplishable task.

Yeah, I’m arguing to not disclose the model. It’s worth far more held close. If you want to work at Jane Street, go work for Jane Street. If you want to build your own models and run your own shop, the tools exist for you to do that without Jane Street (although there’s probably some amount of value learning the ropes there while they pay you, if that’s your thing). My comments in thread are primarily around not havi…

A backtested model is similar to a great startup idea. There is a huge amount of work to be done before it is worth much.

Re: Jane Street Market Prediction ($100k Kaggle competition)

#89

Post by madrafi »

npmisdown wrote: post #79

I'm sorry, if you could build a model to predict markets, why will you post in to Kaggle to get $40k in prize instead of applying this model to your own broker account?

well because quant trading isn't about import xgboost, you need a sustainable infra to handle api failovers, bad data... not even going to mention risk management which is 50% of what quant trading is about. the data provided is anonymized but would probably be a mix of laggard measurements (moving averages, rsi...) and maybe some flow data... quant trading isn't really about finding "secret stuff" most profitable strats you can deploy can be based on stat-arb, basis trading or even just delta-neutral funding farming and such

Re: Jane Street Market Prediction ($100k Kaggle competition)

#90

Post by hogFeast »

riazrizvi wrote: post #51
usmannk wrote: post #24

As a frequent Kaggler (perhaps too frequent... it's a bit addicting, in a way I'm sure others on HN will understand), I was fairly intrigued to see this one pop up in the competition list a few days ago. Finance shops have tried their hand at Kaggle before, but I think they've normally been out of their domain. e.g. Two Sigma recently did a reinforcement learning game competition. I'd caution the HN crowd not to expe…

Mathematical analysis of financial markets is more celebrated when applied to relative valuation of different assets, rather than prediction of the market. Black-scholes, for example, applied calculus with an underlying no-arbitrage assumption to create a thriving market in option pricing, by giving traders a mechanism to reduce risk and thereby reduce bid offer spreads. Same in fixed income, mortgage, and credit mar…

Also, prices aren't stationary. For an equity security, you are predicting the price for a company that is compounding capital over time. You can predict the price for the company at one point, the relative valuation for the company (for example, against peer group) may not change in one year but that company is investing their capital at X% so you get price growth.

The reason why relative valuation models are more effective is the same reason why most sports betting models use current odds as an input. Prices contain information but, in my experience, these methods aren't totally effective because they often miss important information about the company itself (big price moves happen because relative valuations are wrong). Value or quality appears to do fundamental work but is often woefully blind (for example, there are proven accounting issues with value strategies...does your average quant understand this? No. Have they ever read a set of accounts? No. They have no hope. None.)

Just imo, I think quant strategies are almost totally worthless beyond liquidity provision (even a strategy like front-running news in FX...humans do this better, and I know people who are still making tons of money doing this). I think there is massive value in that mode of analysis but the people who make the most are always going to be people who know the fundamentals better (I think firms like Marshall Wace that are doing this synthesis will move ahead) because that information is often not in the price at all.

Post reply on HN