🤖 machine learning

Online Learning of Scale Parameters in Score-Driven Filters

This paper proposes an online learning framework for score-driven filters that treats the gain parameter as a decision variable optimized via a Kullback-Leibler objective, establishing dynamic-regret bounds for mirror-descent updates and demonstrating through simulations and equity-index data that bounded mirror gains effectively adapt to time-varying conditions and outperform constant gains, particularly during market crises.

Original authors: Fabrizio Lillo, Giulia Livieri, Gianluca Palmari

Published 2026-08-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Fabrizio Lillo, Giulia Livieri, Gianluca Palmari

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather, but instead of a single forecast, you have a team of meteorologists who constantly update their models based on the latest temperature readings. In the world of finance and economics, this is similar to how we try to predict things like stock market volatility or interest rates. We use mathematical "filters" that take new data (like a sudden stock price drop) and adjust our predictions for tomorrow. These filters work by looking at a "score"—a signal that tells us how wrong our last guess was—and then taking a step to correct it.

Think of this correction step like walking down a hill to find the lowest point. The "score" tells you which direction is downhill, but you also need to decide how big a step to take. If your steps are too tiny, you'll never reach the bottom; if they are too giant, you might overshoot and fall into a valley on the other side. In traditional models, the size of this step is usually a fixed setting, like a thermostat set to a specific temperature. But what if the weather is changing rapidly? A fixed setting might be too slow to react to a storm or too jumpy during a calm day. This paper asks a simple but powerful question: Can we teach the filter to learn the perfect step size on the fly, adjusting it moment-by-moment as new data arrives?

The authors of this paper, Fabrizio Lillo, Giulia Livieri, and Gianluca Palmari, tackle this problem by treating the "step size" (which they call the gain) not as a static setting, but as a decision variable that can be learned in real-time. They propose a method where the filter looks at the most recent data, calculates a "loss" (how bad the prediction was), and uses a clever mathematical trick called mirror descent to adjust the step size. Imagine a hiker who doesn't just walk in a straight line but uses a special map that warps the terrain to help them stay on the path, even when the ground is slippery or the path is narrow. This "map" is the link function, which ensures the step size stays within safe, logical bounds (it can't be negative or infinitely large).

The paper finds that by using this online learning approach, the filter can adapt much better to changing market conditions than if it used a fixed step size. In their simulations and tests on real stock market data, the "learning" filter generally performed as well as or better than the old methods, especially during times of crisis when the market was swinging wildly. It managed to avoid the extreme, erratic jumps that sometimes happen with other methods, effectively finding a sweet spot between being too cautious and too reckless. The authors show that this works because the "score" from the data actually contains the exact mathematical clue (a gradient) needed to adjust the step size perfectly, provided you use the right geometric map to interpret it.

However, the paper is careful to note that this isn't a magic bullet that solves all prediction problems. The improvements are most visible in specific scenarios, like multi-crisis markets, and the method relies on certain mathematical conditions being met (like the data behaving in a somewhat predictable, convex way). While the theory proves that the method has strong guarantees against "regret" (meaning it won't perform much worse than the best possible step size we could have chosen if we knew the future), the real-world tests are based on simulations and historical data. The authors suggest that while the "bounded" version of their method (which keeps step sizes within safe limits) is very robust, the "unbounded" version (which allows for infinite steps) might be riskier without extra safeguards. Ultimately, this work provides a new, more flexible toolkit for economists and data scientists to build smarter, more responsive prediction models that can dance with the market rather than just marching to a fixed beat.

Technical Summary: Online Learning of Scale Parameters in Score-Driven Filters

Problem Formulation
Score-driven filters are observation-driven time-varying parameter models where the state update direction is determined by the score of the conditional log-likelihood, scaled by a local information or curvature matrix. While the direction of the update is well-understood, the scale parameter (referred to as the gain) is typically treated as a static parameter or, in accelerated score-driven models (aGAS), as a time-varying variable driven by score products without a unified predictive objective.

This paper reframes the gain selection problem as an online predictive decision problem. The authors treat the gain not as a structural latent parameter but as a decision variable chosen after the current observation and scaled score are revealed, but before the next observation occurs. The core problem is to select a gain from an admissible set that minimizes the conditional one-step-ahead predictive loss (negative log-likelihood) or, equivalently, the Kullback–Leibler (KL) divergence to the true conditional density, restricted to the set of states reachable via the realized score direction.

Methodology
The authors develop a framework integrating online convex optimization with econometric filtering theory:

Conditional Reachable Sets: Once the scaled score $d_t$ is realized, the admissible gains map to a "reachable set" of next states. For a scalar gain, this is a line segment; for a diagonal gain, it is a coordinate-aligned box. The gain selection is thus a choice of a point within this score-conditioned set.
Stochastic Gradient Identification: In the scalar unscaled case, the authors prove that the negative raw product of consecutive scores is the stochastic gradient of the predictive gain loss. This establishes that the feedback mechanism used in aGAS is not merely a heuristic for serial dependence but an exact gradient descent signal for the predictive objective.
Mirror Descent Geometry: To handle constraints (e.g., positivity, boundedness) on the gain, the authors employ Online Mirror Descent (OMD). The link function used to parametrize the gain (e.g., logistic, exponential) acts as a mirror map, inducing a non-Euclidean geometry (via Bregman divergence) on the gain domain. This geometry dictates how the gain updates in response to the gradient signal.
Persistence as Bregman Pull-back: The persistence mechanism found in empirical aGAS specifications is interpreted as a "pull-back" towards a reference gain within the Bregman geometry. The intercept in the latent recursion determines the reference gain level, while the persistence parameter controls the memory (discounting) of past gradient signals.
Dynamic Regret Analysis: The paper establishes dynamic-regret bounds for the adaptive gain relative to a time-varying, $F_t$ -measurable comparator sequence (the "reachable oracle"). These bounds hold under convexity, compactness, and regularity conditions, comparing the adaptive gain against the best possible gain choice given the realized score at each step.

Key Contributions
The paper makes five specific contributions:

Conditional KL Objective: It defines a conditional gain objective where minimizing the predictive loss is equivalent to finding the KL-projection of the true density onto the family of densities reachable through the realized score direction.
Gradient Identity: It identifies the product of consecutive scores as the stochastic gradient of this predictive loss, providing a variational justification for score-product feedback.
Link as Mirror Map: It interprets the link function not just as a constraint enforcer but as the geometric structure (mirror map) governing the update dynamics.
Persistence Interpretation: It provides a geometric interpretation of persistence and intercepts in aGAS as Bregman pull-backs towards a reference gain, separating the roles of memory, reference level, and learning rate.
Regret Guarantees: It derives dynamic-regret bounds for bounded mirror updates (projected and discounted) against moving comparators. These bounds explicitly quantify the trade-off between adaptivity (tracking a moving target) and stability (bias towards a reference gain).

Results

Theoretical: The authors prove that under convexity of the pulled-back loss, the adaptive gain sequence satisfies a dynamic-regret bound. The bound consists of initialization costs, a tracking cost dependent on the variation of the comparator path, and a gradient energy term. For persistent recursions, an explicit bias term appears, measuring the cost of shrinking towards the reference gain.
Numerical: Synthetic experiments isolate the mechanisms of the framework. They demonstrate that:
- Adaptive gains successfully track moving gain oracles.
- Scaling rules significantly impact the geometry of the reachable set and the learnability of gains.
- The dynamic-regret bound accurately reflects the cost of tracking rapidly changing comparators.
- Bounded mirror gains (e.g., logistic link) generally match or outperform constant gains and avoid the extreme spikes associated with unbounded exponential links, particularly in multi-crisis market scenarios.
Empirical: An out-of-sample panel of equity-index volatilities shows that bounded mirror gains perform robustly, avoiding the instability of unbounded specifications while adapting effectively to structural breaks.

Significance and Claims
The paper claims to provide a rigorous econometric foundation for the adaptive gain in observation-driven filters. By treating the gain as an online learning variable with a specific predictive objective, geometry, and benchmark, the authors move beyond the "acceleration" interpretation of aGAS.

The significance lies in:

Unification: Connecting score-driven filtering with online convex optimization and mirror descent.
Interpretability: Offering a clear geometric and variational interpretation of existing empirical specifications (like aGAS), showing them to be instances of discounted mirror descent with specific link functions.
Theoretical Guarantees: Providing the first dynamic-regret guarantees for the gain component of score-driven filters, conditional on the realized score path. This isolates the learning of the gain from the stability of the state path and the optimality of the score direction.
Practical Guidance: Suggesting that bounded link geometries (e.g., logistic) offer a superior trade-off between adaptivity and stability compared to unbounded exponential links, particularly in volatile or non-stationary environments.

The authors explicitly note that their regret guarantees are conditional on the realized score sequence and do not constitute a full policy-regret theorem for the entire state recursion, as different gain histories would generate different future contexts. The analysis focuses on the local learning mechanism of the gain layer.

More like this