| Why do algorithmic trading articles pop up frequently on HN? I understand why people are attracted to it because they think that they can beat the market by implementing their own stat arb strategy using neural networks/genetic programming etc. They think that they'll fare out better than the average day-traders because they have the know-how's to program and the quantitative skills to formulate the models and the technical skills to implement the models. But what people don't realize is that despite of all that, you still can't beat the street because: speed and volume. All high frequency trading strategy rely on fast execution speed (get in and get out, most prop shops at investment banks and hedge funds hold their positions usually for a couple of seconds, or even less), because due to prevalence of electronic trading, the inefficiency in the market for arbitrage opportunities exists in the time-frame of seconds or micro-seconds. In arbitrage, it's no longer a matter of who can discover the treasure map; it's more like everyone already has the treasure map, whoever can get to the X mark the spot first wins. Goldman Sachs/SAC/DE Shaw etc. all "colo" with the market exchanges, meaning they have their algorithmic trading servers hooked up directly to the same trading servers of NYSE/Nasdaq/BATS. They might be connected directly to those market servers in the same server room in NYC/Kansas, or they might be connected over a dark fiber network in Jersey City. In the matter of seconds a retail brokerage customer puts out a buy order for GOOG, or even a aspiring trading-hacker sends out a order over Interactive Broker API, these algo machines have already piped out thousands of orders and a good percentage of those orders have already been filled. Secondly, good algo's require fast real-time information processing of the quote book, general market condition, real-time calculation of the basket of related stocks in the stat arb strategy. On an single active stock, there might be hundreds of bid/ask/trade ticks per minute from a single market center. Can an individual's machine setup handle real-time analysis of a basket of stocks and general market condition from multiple market centers? This issue is so critical and complex that there exists a sub-industry (Complex Event Processing, CEP) and technology companies (StreamBase) that created and profit from dealing with this problem for hedge funds and prop shops. Finally, usually high frequency algo's are only profitable with high volume and low transaction cost. Typically, arbitrageurs and liquidity providers make pennies on the share per trade. But their daily volume are in the tens or hundreds of millions shares traded; their transaction cost are in the quarters of pennies or sometimes negative (meaning they are compensated by the exchanges for providing liquidity). Compare this to a retail consumer account with $10,000 and trading comission of $5 (ETrade) - $0.01/share (IB). Still, I know of people who have IB API setup's and profit consistently. But I wanted to let would-be traders know the competitive disadvantage that they are up against the big boys before they pour out their life savings in their neural network futures arbitrage algorithm. |