4 ms·

A lot of the high frequency trading code actually is very expensive to recreate, because it contains logic like (and I am paraphrasing here): if(bookPressure >

A lot of the high frequency trading code actually is very expensive to recreate, because it contains logic like (and I am paraphrasing here): if(bookPressure > 0.7 and alpha23signal<0.2) buyEurUsd() where constants like 0.7 and 0.2 were arrived at from looking at actual trade activity (and order book fill rates) conducted by the firm over a period of a few months. That can be incredibly expensive and or impossible to "recreate" from scratch. The exact original code is also useful if you wanted to write code to "detect" that some automated trading is being done by a Goldmans proprietary algorithm, and then anticipate what it's going to do next so you can trade ahead of it.

tptacek 11y ago
This is an interesting point I hadn't seen made before. If it applies here, then what you're saying is that the structure of the code itself doesn't matter, so much as the parameters encoded into it --- those parameters being available only to people who can bankroll the trading desks that generate the information from which they're derived. In that scenario, the "code" itself (as we understand "code") is a red herring, right? I'm surprised if this is what actually happened; it's too reasonable-sounding.
lordnacho 11y ago
Right, so that's a parameter, isn't it? A mathematical number that fits into a model. In which case you need to understand the model.
- Patient0 11y ago
  Right but I don't see how this detracts from my point. Also, often the models are very simple - it's the calibration to a particular market that is very expensive (the constants - which are then often hard coded for best performance). This is because calibrating it requires access to data from actual trading activity (not just tick data, I mean also information on say the chance that your order will be filled in a particular situation: information you can only get by actually placing orders of size in the market and observing when they got filled).
  - lordnacho 11y ago
    If you don't know how that constant is arrived at, how do you know whether it still applies to the current market? Don't you want to know whether your assumptions still hold? And aren't you stealing the wrong thing, then? If you do know, why do you care what the number actually is? And why would hard coding improve performance? Compiler optimization?
  - dropit_sphere 11y ago
    >(the constants - which are then often hard coded for best performance). This caught my eye. Really?
    - tedunangst 11y ago
      #define TARGET_DELTA 70 Is a reasonable form of "hard coding".
malandrew 11y ago
Ewwww... magic numbers. I would hope that I would encounter source code that at least took me to the code that derived these magic numbers. Why 0.7 for bookPressure or 0.2 for alpha23signal?

tptacek 11y ago

This is an interesting point I hadn't seen made before. If it applies here, then what you're saying is that the structure of the code itself doesn't matter, so much as the parameters encoded into it --- those parameters being available only to people who can bankroll the trading desks that generate the information from which they're derived. In that scenario, the "code" itself (as we understand "code") is a red herring, right? I'm surprised if this is what actually happened; it's too reasonable-sounding.

lordnacho 11y ago

Right, so that's a parameter, isn't it? A mathematical number that fits into a model. In which case you need to understand the model.

Patient0 11y ago
Right but I don't see how this detracts from my point. Also, often the models are very simple - it's the calibration to a particular market that is very expensive (the constants - which are then often hard coded for best performance). This is because calibrating it requires access to data from actual trading activity (not just tick data, I mean also information on say the chance that your order will be filled in a particular situation: information you can only get by actually placing orders of size in the market and observing when they got filled).
- lordnacho 11y ago
  If you don't know how that constant is arrived at, how do you know whether it still applies to the current market? Don't you want to know whether your assumptions still hold? And aren't you stealing the wrong thing, then? If you do know, why do you care what the number actually is? And why would hard coding improve performance? Compiler optimization?
- dropit_sphere 11y ago
  >(the constants - which are then often hard coded for best performance). This caught my eye. Really?
  - tedunangst 11y ago
    #define TARGET_DELTA 70 Is a reasonable form of "hard coding".

malandrew 11y ago

Ewwww... magic numbers. I would hope that I would encounter source code that at least took me to the code that derived these magic numbers. Why 0.7 for bookPressure or 0.2 for alpha23signal?