Feature Extraction Using Alternative Data

Alternative data is very useful because it can be used to extract features, which are then used as inputs to other models that predict:

• Returns of stocks: These predicted returns can then be used as inputs in portfolio construction, such as for long/short strategies.

  For this, we can use a mean-variance optimization framework.

  Here we can:

  • Maximize the Sharpe ratio.

  • Maximize expected returns minus risk minus transaction costs.

  • Minimize portfolio variance subject to a target return.

  Subject to constraints such as:

  • Beta-neutral constraint, \( \beta_p = \sum_i w_i \beta_i = 0 \). This is especially important for hedge funds because hedge funds are mostly paid to generate returns that are not correlated with the market.

  • Target return constraint, \( \sum_i w_i \mu_i \geq \mu^* \).

  • Dollar-neutral constraint, where weights sum to zero, \( \sum_i w_i = 0 \), so dollars long equal dollars short.

  • Weight limits, \( l_i \leq w_i \leq u_i \), so that no single position becomes too large.

  • Leverage constraint, so that the sum of absolute weights, \( \sum_i |w_i| \leq L \), does not exceed some limit.

• Volatility: Alternative-data features can also be used to predict future realized volatility. We can then compare forecasted realized volatility with implied volatility and use this relationship to buy or sell volatility using options. Two ways to do this are:

  • Long or short straddles.

  • Delta-neutral strategies where we are long or short gamma.

  If forecasted realized volatility is higher than implied volatility, we may want to be long volatility, for example by buying straddles or taking delta-hedged long-gamma positions. If forecasted realized volatility is lower than implied volatility, we may want to be short volatility, for example by selling straddles or taking delta-hedged short-gamma positions.