A Computational Linguist on Wall Street

Discussion in 'Philosophy and Strategy' started by itfin, Sep 4, 2007.

  1. Conclusions

    It is becoming increasingly harder to find times series patterns with enough predictive power to profit from them: as market participants discover and exploit them, they thereby eradicate them.

    In the seventies, most trend following sstem were profitable, as treng reversals were relatively easy to spot.

    Over time, more and more sophisticated statistical pattern recognition capabilities were needed, due to -- and the cause of -- less predictability and reduced profits, feeding a vicious cycle.

    I've repeatedly observed in historical data that some trading stratedy is profitable with a high Sharpe ratio until a certain point in time, after which it ceases to be profitable.

    This point in time is typically earlier when trading equities than currencies, and later still when trading (Treasury) bonds.

    Conclusions, cont'd

    It is better to learn trading strategies than market prediction. Good prediction can still result in poor trading -- you typically gain a little when you're right and lose a lot when you're wrong.

    One must use increasingly more complex statistical pattern discovery techniques to trade profitably based on time series analysis, and the margins grow increasingly thinner.

    It's time for new approaches, e.g., to feed live newswire to the traders, employing advanced NLP and automated reasoning.

    The poor quality of historical data is a serious problem. It hampers practitioners and call into question much academic research.

    The inverse 'survivor bias' applies: why publish positive predictability results when one can instead exploit them for personal gains?