Sequential Tracking in Pricing Financial Options using Model Based and Neural Network Approaches

Mahesan NiranjanNeurIPS 1996

TL;DR

This paper demonstrates how the Extended Kalman Filter (EKF) can be used to track financial option prices in real-time by treating model parameters as hidden states. By applying this sequential estimation to both the classic Black-Scholes model and Radial Basis Function (RBF) neural networks, the author shows that "tracking" the market is significantly more accurate than using traditional historical averages or batch-processed data.

What Problem Does This Paper Address?

Pricing financial options is notoriously difficult because the underlying markets are "non-stationary"—meaning the statistical properties of the system change over time. Traditional finance models, like Black-Scholes, often rely on parameters (such as volatility) estimated from a "moving window" of past data. This creates a lag; by the time you've calculated the volatility of the last 50 days, the market has already moved on.

The author argues that because financial data arrives sequentially, our models should update sequentially. The challenge is to find a way to update complex, nonlinear models (both theoretical and neural) every time a new price tick occurs, ensuring the model "tracks" the current state of the market rather than just averaging its history.

Methods

The core engine of this paper is the Extended Kalman Filter (EKF), specifically a second-order version that uses Taylor series expansions to handle nonlinearities. The author applies the EKF to two different modeling paradigms:

  1. Model-Based (Black-Scholes): The author treats the Black-Scholes formula as the observation equation. The "hidden states" being tracked are the volatility ($\sigma$) and the risk-free interest rate ($r$). By calculating the Jacobian and Hessian matrices for these variables, the EKF can update its estimate of $\sigma$ and $r$ with every new market observation.
  2. Neural Network (Radial Basis Functions): The author uses an RBF network, a type of neural network that excels at function approximation. While the nonlinear centers of the RBF are fixed, the linear weights are updated using the Kalman filter. This allows the neural network to adapt its "shape" to the market's changing behavior in real-time.

The paper also compares these to "Trivial" benchmarks (using the previous day's price) and "Historic" benchmarks (using standard moving-window estimates).

Datasets

The study uses real-world market data:

  • FTSE-100 Index Options: Data spanning February 1994 to December 1994.
  • Specifics: The author focuses on five pairs of Call and Put options with identical strike prices (2925, 3025, 3125, 3225, and 3325) and a shared maturity date of December 1994.

Key Insights and Ideas

  • Implied Interest Rates: In standard finance, the interest rate is usually taken from bond market data. Here, the author allows the EKF to estimate an "implied interest rate" that best fits the option prices. The author suggests that the gap between this "implied" rate and the "market" rate might actually represent trader speculation or currency risk.
  • Sequential > Batch: The results clearly show that updating a model one step at a time (Sequential) is far superior to re-estimating a model using a block of data (Batch). The sequential approach allows the model to "forget" old, irrelevant data more effectively.
  • Hybrid Strength: The paper shows that the EKF is a versatile tool that can bridge the gap between "physics-based" financial theory (Black-Scholes) and "data-driven" machine learning (RBF networks).

Key Results

The quantitative results (measured in squared approximation errors) show a clear victory for the Kalman-tracked models:

  • Black-Scholes (BS): At a strike price of 3125, the BS Kalman approach achieved an error of 0.0112, compared to 0.0343 for the standard BS Historic method.
  • Neural Networks (RBF): For the same strike price, the RBF Kalman achieved an error of 0.0193, significantly outperforming the RBF Batch error of 0.0455.
  • Consistency: Across all tested strike prices, the Kalman-based versions of both models consistently produced the lowest errors, often by a factor of 2 or 3 compared to their non-sequential counterparts.

Limitations and Future Work

  • The "Greeks": The author notes that while the RBF model tracks prices well, it wasn't specifically trained to track derivatives (like Delta, which measures sensitivity to price changes). In real-world trading, tracking Delta is essential for hedging.
  • Parameter Tuning: The noise covariance matrices ($Q$ and $R$) in the Kalman filter were set by trial and error. The author suggests that more systematic, multi-model approaches could automate this tuning.
  • Future Direction: A promising next step suggested is to train neural networks to approximate both the option price and its derivative simultaneously, creating a "delta-neutral" tracking system that would be more practical for professional traders.

Summary updated

BibTeX
@inproceedings{niranjan1996sequential,
  author = {Mahesan Niranjan},
  title = {{Sequential Tracking in Pricing Financial Options using Model Based and Neural Network Approaches}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1996},
}
Figure