📈 economics

Clustered Local Projections for Short and Ultra-Short Time Series -- A Hierarchical Bayesian Framework

This paper introduces a hierarchical Bayesian framework with sparse finite mixture clustering to improve local projection impulse response estimation for short and unbalanced time series by borrowing strength across related units, demonstrating its effectiveness through simulations and empirical analysis of heterogeneous price reactions to supply-chain and oil shocks.

Original authors: Todd Clark, Florian Huber

Published 2026-08-06

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Todd Clark, Florian Huber

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 a detective trying to solve a mystery, but you only have a tiny, blurry snapshot of the crime scene instead of a full video. In the world of economics, this is a common problem. Economists use tools called "Local Projections" to figure out how the economy reacts to big shocks, like a sudden oil shortage or a supply chain breakdown. They want to know: "If we hit this button today, what happens to prices next month? Or next year?" The trouble is, these tools are data-hungry. They need long histories of numbers to work well. But sometimes, economists only have a few months of data for a new type of price index, or the time they are looking at is so far in the future that their tiny dataset runs out of steam before the answer arrives. It's like trying to predict the weather for next December based on only three days of temperature readings.

To fix this, the authors of this paper, Todd E. Clark and Florian Huber, invented a clever way to let the short stories borrow from the long ones. They built a "hierarchical Bayesian framework," which sounds fancy but is basically a smart system for grouping similar things together. Think of it like a classroom where the teacher asks a question. If a student with only a few notes (a short time series) is asked to answer, they might be stuck. But if the teacher knows that this student is in the same "study group" as a student with a thick textbook (a long time series), the teacher can let them share the answer. The key is knowing who belongs in which group. If you mix up a group of students who love math with a group who love art, the math student won't get the right help. This paper figures out how to automatically sort these economic time series into the right groups based on how they actually behave, rather than just guessing.

The Problem: The "Short Memory" of Economics

Economists often study how the economy reacts to shocks, like a sudden spike in oil prices or a disruption in shipping. They use a method called Local Projections (LPs). Imagine you want to see how a price changes over time after a shock. You run a separate math equation for every single step into the future (one month later, two months later, etc.). This is great because it's flexible and doesn't force the economy into a rigid box. However, there's a catch: every step into the future eats up a chunk of your data. If you have a time series that only goes back 50 months, you can't really say much about what happens 40 months from now because you've run out of history. It's like trying to guess the ending of a movie when you've only seen the first five minutes.

In the real world, we have many new economic indicators that are very short. Maybe a new survey started last year, or a specific type of price index was only recently created. These "ultra-short" series are useless for looking far into the future using standard methods. The authors noticed that while these short series are new, they are often related to older, longer series. For example, a new "Services Producer Price Index" might react to an oil shock in a very similar way to an old "Manufacturing Price Index" that has data going back to the 1950s. The question is: How do we use the long history of the old series to help us understand the short history of the new one, without forcing them to be exactly the same?

The Solution: A Smart Grouping System

The authors propose a solution that acts like a super-smart librarian. Instead of treating every time series as a lonely island, their method groups them into "clusters" based on how they actually behave.

Here is how the magic happens:

The Borrowing Trick: The system assumes that if two series react similarly to a shock, they should help each other. If a short series has almost no data for a specific future month, the system looks at the long series in its group and says, "Okay, since you have no data, let's use the pattern from your long-term friends." This allows the short series to "borrow" information to make predictions far into the future that it couldn't reach on its own.
The "One Size Fits All" Trap: The authors warn that you can't just lump everyone together. Gasoline prices and hospital service prices don't react the same way to a shock. If you force them into one big group, you get the wrong answer. So, their method uses a sparse finite mixture. This is a fancy way of saying the system automatically figures out how many groups there are and which series belong in which. It's like a party where the guests naturally split into different circles of friends based on their interests, rather than being forced into one big circle.
The Three-Layer Safety Net: The system has three layers of organization. First, it groups similar series together. Second, it realizes that even small groups might need a little extra help, so it connects those groups to a "population center" (the average of everyone). This ensures that even a group with just one member can still learn something from the whole crowd, just a little less intensely.

What They Found: Better Answers for Short Data

The authors tested their idea using simulations based on real US economic data. They created a fake world with 80 different time series: half were long and healthy, and half were very short (some with as few as 25 months of data). They then asked: "Who can predict the future best?"

The results were clear:

For the short series: The new method was a game-changer. It reduced the error in predictions by about 50% to 90% compared to the old standard method. In some cases, it cut the error in half at short, medium, and long horizons. It was so much better that the old method looked almost useless for the shortest series.
For the long series: The new method didn't hurt the long series at all. It produced results almost identical to the standard method, proving that borrowing from others doesn't mess up the people who already have plenty of data.
The Importance of Grouping: They found that if you just forced everyone into one single group (ignoring differences), the results got worse. The "smart grouping" was the secret sauce. It allowed the system to say, "These two are alike, so let's share," while also saying, "These two are different, so let's keep them separate."

The Real-World Test: Prices and Shocks

To see if this worked in the real world, the authors applied their method to 43 different price series in the US. This included old, well-known price indices (some going back to 1959) and newer, shorter surveys from Federal Reserve banks. They looked at how these prices reacted to two specific shocks: a supply-chain shock (like shipping delays) and an oil supply shock.

The method sorted the 43 series into 6 or 7 distinct clusters.

Cluster 1 was a big group of 20 series, mostly consumer prices and some producer prices.
Cluster 4 was a group of 6 regional surveys.
Some series, like the price of final demand goods, were so unique they formed their own single-member cluster.

The results showed interesting patterns:

Headline vs. Core: When a shock hit, the "headline" prices (which include food and energy) reacted much more sharply than the "core" prices (which exclude them).
Goods vs. Services: Goods prices changed more than services prices.
The Unprocessed Effect: The prices of unprocessed goods (the raw materials) moved the most, which makes sense because they are the first to feel the shock.

Crucially, the method gave much sharper, more precise answers for the newer, shorter series. For example, some of the newer service price surveys only had about 33 usable data points at the longest time horizon. Without this method, economists would have had to guess wildly. With the method, they could see clear patterns because the short series borrowed strength from the longer ones in their cluster.

Why It Matters

This paper doesn't just offer a new math trick; it offers a way to see the future more clearly when the past is foggy. By letting short time series borrow wisdom from long ones, but only when they are truly similar, the authors have created a tool that makes economic forecasting more accurate for the newest and most fragile data. It suggests that in a world of changing data, we don't have to throw away the short stories; we just need to find the right friends to help us read them.

Technical Summary: Clustered Local Projections for Short and Ultra-Short Time Series

1. The Problem
Local Projections (LPs), introduced by Jordà (2005), are a standard tool for estimating the dynamic causal effects of structural shocks. However, their application to short and ultra-short time series is severely hampered by a lack of degrees of freedom. In macroeconomic settings, samples are often short, and newly available indicators (e.g., specific producer price indices or regional survey data) may have sample lengths so small that the horizon of interest ( $h$ ) exceeds the available data ( $T$ ). Standard Ordinary Least Squares (OLS) estimation in these contexts yields unreliable inference, with coverage rates falling below nominal levels and estimates becoming impossible to compute when $h > T$ . Furthermore, many macroeconomic panels are unbalanced, containing a mix of long-established series and short, newly introduced ones.

2. Methodology
The authors propose a Bayesian hierarchical framework designed to estimate LP impulse response functions (IRFs) across a panel of related time series with heterogeneous sample lengths. The methodology consists of four key components:

Hierarchical Pooling: Instead of estimating each series independently, the framework pools information across the panel. The response coefficients ( $\rho_{i,h}$ ) for series $i$ at horizon $h$ are modeled as draws from a common distribution. The posterior mean for any series becomes a precision-weighted average of its own OLS estimate and the panel-wide (or cluster-specific) mean. As a series' sample size shrinks, the weight on the pooled prior increases endogenously; if a horizon exceeds the series' effective sample length, the estimate relies entirely on the pooled information.
Sparse Finite Mixtures: Recognizing that macroeconomic series often exhibit heterogeneous dynamics (e.g., gasoline prices vs. hospital services reacting differently to shocks), the authors generalize the pooling prior to a sparse finite mixture. Series are probabilistically assigned to a small number of latent clusters based on the similarity of their IRF profiles. The effective number of clusters is determined endogenously via a sparse Dirichlet prior on the mixture weights, allowing the model to "prune" unnecessary components.
Three-Level Hierarchy: To handle small clusters (including singletons) efficiently, the model employs a three-level hierarchy. Cluster means are themselves drawn from a common population center. This ensures that even a cluster with a single member can borrow strength from the broader panel, albeit more weakly than larger clusters, preventing the model from reverting to a purely series-specific estimate when cross-sectional information is sparse.
Robustification of Uncertainty: LP errors possess a moving average (MA) structure, causing standard posterior credible sets to understate uncertainty. Following Müller (2013), the authors rescale posterior draws ex post to achieve correct frequentist coverage. Crucially, they pool the long-run variance estimates at the cluster level rather than the unit level. This is essential because unit-specific variance estimates are extremely noisy in short samples; pooling stabilizes the correction.

The framework accommodates unbalanced panels without interpolation or data balancing. Short series simply contribute fewer observations to the cross-sectional updating steps, and the model naturally imputes responses for horizons beyond the series' own data length.

3. Simulation Results
The authors evaluate the framework using a simulation calibrated to US macroeconomic data (FRED-MD), comparing their proposed estimators against standard series-by-series OLS LPs.

Point Estimation Accuracy: For short and ultra-short series, the proposed mixture-based pooling approach substantially reduces the Mean Absolute Error (MAE) of IRF estimates compared to the standard LP benchmark. In the "very short" regime ( $T \in [25, 60]$ ), the full model reduces MAE by approximately 80–90% across horizons. For long series ( $T=500$ ), the pooled estimates track the OLS benchmark closely, indicating that pooling does not introduce significant bias when the own-data signal is strong.
Bias-Variance Trade-off: The pooling approach significantly reduces estimator variance. While single-cluster pooling (forcing a common response) can introduce bias if series are heterogeneous, the sparse finite mixture adapts to the data, isolating heterogeneous series into their own components to avoid spurious precision.
Coverage Rates: Standard LP confidence intervals for short series suffer from severe under-coverage (dropping to ~25–50% at long horizons). The proposed cluster-pooled correction restores coverage to near-nominal levels (90% or slightly higher), whereas unit-specific corrections remain noisy and ineffective in short samples.

4. Empirical Application
The framework is applied to a panel of 43 US price series (33 consumer/producer price indices and 10 regional Federal Reserve business survey measures) to study responses to supply-chain and oil supply shocks.

Clustering: The model identifies 6–7 distinct clusters based on IRF similarity, grouping long-running aggregates with shorter, newer series (e.g., newer services PPIs).
Heterogeneous Responses: The results reveal that supply-chain and oil shocks trigger heterogeneous reactions. Headline price indices respond more sharply than core counterparts, and goods prices change more than services prices. Specifically, unprocessed goods for intermediate demand show the sharpest responses.
Precision: The cluster-mean IRFs are estimated with significantly higher precision (tighter credible bands) than individual series estimates, particularly for the short series that benefit from borrowing strength from longer, similar series.

5. Significance and Claims
The paper claims to offer a solution to the "data-hungry" nature of LPs in the context of short and unbalanced macroeconomic panels. Its primary contributions are:

Methodological: A hierarchical Bayesian framework that explicitly handles unbalanced panels and ultra-short series by borrowing information across related time series without requiring data balancing.
Flexibility: The use of sparse finite mixtures allows the model to accommodate heterogeneous dynamics, avoiding the bias of forcing a single common response while still leveraging cross-sectional information.
Practical Utility: The approach enables the estimation of IRFs at horizons that exceed the length of individual short series, a task impossible with standard methods.
Empirical Insight: The application demonstrates that standard methods often fail to utilize newer, shorter price series effectively, whereas the proposed framework can integrate them to reveal nuanced, heterogeneous responses to structural shocks.

The authors conclude that while the gains are concentrated in short samples, the approach produces estimates for long series that are comparable to current best-practice benchmarks, suggesting a robust tool for modern macroeconomic analysis where data availability is increasingly heterogeneous.

The Problem: The "Short Memory" of Economics

The Solution: A Smart Grouping System

What They Found: Better Answers for Short Data

The Real-World Test: Prices and Shocks

Why It Matters

More like this