3 ms·

Trading algos are trained on historical market data, this can (and should) include crashes etc.

• dnautics 9y ago
  Obviously, but 1) the data are not the same then so it's possible the algos just learn that featureset and not the actual patterns, and 2) I'm more talking about the meta-pruning and algorithmic balancing quantopian uses to determine which bot wins, which fundamentally can't be trained on the past.
• justrobert 9y ago
  Back testing based on OHLC on a crash day isn't very useful for many systems. The bid/ask of instruments become illogical. Bad trade data in the live feed won't be in your historical data. Intraday during a liquidity drop is mostly based on your relationship with your broker. How they handle margin calls will decide if your trades survive. For example, traders have had margin calls on option spreads with a known fixed risk at the time of opening. But the individual legs of the trades were making very incorrect markets, so the broker algorithms issued margin calls and liquidated at the worst time possible.