> Thanks to increases in computing power, Sentient can squeeze 1,800 simulated trading days into a few minutes.
How is it possible to generate seven years of convincing sample data from historical trading data without overfitting your models?
How is it possible to generate seven years of convincing sample data from historical trading data without overfitting your models?
From the article:
> These genes are tested by giving them hypothetical sums of money to trade in simulated situations created from historical data.
So they simply use historical data (going 7 years back). This is fed into a simulator (a trading day where you have access to the data of all days leading up to the simulated trading day).
> without overfitting your models?
Traditionally backtesting is used: http://www.investopedia.com/terms/b/backtesting.asp
Care should be taken to run the tests over significant periods of time (to reflect changing market conditions) and to have a final out-of-time holdout set to lessen the effects of picking "winners" that were winning purely due to chance (introduced by cherry-picking winners from thousands of models).
Simple -- create a large set of models, each using different selected parameters, run them against the market, and pick the outcomes that make the scheme look good. It's called "data mining."
A famous "psychic," who shall go nameless, made a career of appearing with a sealed, dated, registered letter, opening it, and proving that she had correctly predicted the outcome of an election / horse race / other event in advance. She was always right, and the registered letters were real and were mailed before the event to be predicted. How did she do it? She mailed herself more than one registered letter for each event. People are sooo ... credulous.