关于Black-Scholes方程模型在资产定价中的适用性及相关问题的技术问询

Hey there, let's unpack your questions about Black-Scholes and its place in modern finance—this is a classic gap between textbook theory and real-world practice that a lot of folks stumble on!

First off, your friend's comment about Black-Scholes being less widely relied on these days is spot-on, and it all boils down to the model's core assumptions clashing with reality. Let's break that down:

• Black-Scholes is built on a set of idealized assumptions: constant volatility, fixed risk-free interest rates, no transaction costs, and asset prices following a smooth geometric Brownian motion (no sudden jumps). None of these hold in real markets. Volatility swings wildly (just look at the VIX index during crises), interest rates shift, and black swan events (like the 2008 crash or 2020 COVID meltdown) cause massive price jumps that the model can't account for.
• The "lost money" part comes from cases where traders relied strictly on Black-Scholes without accounting for these flaws. A famous example is Long-Term Capital Management (LTCM), a hedge fund run by Nobel Prize-winning economists that used SDE-based models but ignored extreme tail risk—they nearly collapsed the global financial system in 1998 when Russia defaulted on its debt, an event their models didn't predict. Even in day-to-day options trading, Black-Scholes prices often deviate sharply from market prices, especially for out-of-the-money options (this is called the "volatility smile" and it's a direct sign the model's constant volatility assumption is wrong).

On your question about fitting SDE models to data:

• You're right that practitioners calibrate parameters like volatility and drift using market data, usually via optimization (like minimizing the difference between model-predicted option prices and actual market quotes). But the problem isn't the optimization method—it's that the model's structure can't capture real-world price dynamics. When you calibrate Black-Scholes, you don't get a single constant volatility; you get a different value for every option strike price and expiration date (the "volatility surface"). This means the model is essentially being forced to fit data it wasn't designed for, leading to inconsistent predictions.
• The error you're thinking of is twofold: first, the model's predictions are often far off from actual prices, especially during volatile periods. Second, the error isn't stable—it changes as market conditions shift, because the model can't adapt to changing volatility or jump risks.

As for alternatives to Black-Scholes:

• Modified SDE models: Many practitioners use extended SDE frameworks that fix key flaws. The Heston model adds stochastic volatility (volatility itself changes over time), while the Merton jump-diffusion model includes sudden price jumps. These models can better fit the volatility smile and account for extreme events.
• Statistical time series tools: The ARIMA, GARCH, ETS models your friend mentioned are actually critical here. GARCH, for example, is widely used to model time-varying volatility, which feeds into more accurate SDE calibrations. These time series methods are often used for short-term price forecasting and volatility estimation, complementing more complex pricing models.
• Machine learning models: Modern hedge funds and quant firms increasingly use ML models (like LSTMs, Transformers, or gradient-boosted trees) to capture nonlinear price relationships that SDEs can't. These models are trained on massive datasets (including high-frequency trading data) to predict price movements or volatility without relying on strict mathematical assumptions.

To circle back to your friend's students: their focus on time series libraries makes perfect sense. A lot of quantitative finance work starts with statistical exploration of data (using those time series tools) before moving to more complex pricing models. It's a practical approach that bridges theory and real-world data.

备注：内容来源于stack exchange，提问作者krishnab