The ad hoc Black-Scholes model is one of the most widely used models for forecasting implied volatility. In this paper, we propose a methodology that provides more accurate out-of-sample implied volatility forecasts. Standard approaches estimate the whole volatility smile using both out-of-the-money puts and calls. The improvements from our method are obtained by taking advantage of information contained in the asymmetric slopes of the put and call implied volatility sneers that result in a discontinuity when moneyness is equal to 1. These improvements in out-of-sample implied volatility forecasts are large and significant. Our results are robust across several dimensions, including: time period, forecast horizon, moneyness, and model specification.