Coherent forecast combination for linearly constrained multiple time series

summary

The gist

Linearly constrained multiple time series, encountered in contexts like National Accounts or multilevel frameworks…

In short

The paper develops an optimization method to combine multiple unbiased base forecasts for linearly constrained time series, like national accounts data. It derives a closed-form solution for the optimal coherent forecast and its error covariance matrix. Empirical testing on Australian electricity generation shows this method consistently outperforms individual forecasts and other combination techniques.

Key concepts

Coherent Forecast Combination
This is the process of merging several separate, potentially conflicting base forecasts into one single, accurate final forecast. The goal is to create a result that respects the underlying constraints of the time series data while minimizing prediction errors.
Linearly Constrained Multiple Time Series
These are multiple time series (like different economic indicators) that must follow specific linear relationships or constraints. This structure is common in complex systems like national accounts, where one variable's forecast depends on others.
Optimal Unbiased Estimator
The study frames the problem as finding the best possible way to combine all available base forecasts into a single estimate that remains unbiased. The authors conclude that this optimal combination method, called occbe, is the most effective overall approach found.

Terminology used across episodes

This episode discusses

The paper

Coherent forecast combination for linearly constrained multiple time series · Read on arXiv

University of Padua

Forecasting linearly constrained multiple time series, as those arising in National Accounts, energy systems and supply chain planning, requires that the forecasts of all variables jointly satisfy the underlying accounting or resource-balance identities. When several experts provide forecasts for some or all of the variables, classical forecast combination operates on each variable and ignores the linear constraints, while forecast reconciliation enforces such constraints by using a single forecast per variable. To bridge this gap, we develop an original coherent forecast combination framework, in which the joint problem of combination across experts and coherence across variables is formulated as a linearly constrained optimization problem. Closed-form expressions are derived for the coherent combined forecast and its error covariance matrix, and the statistical properties of the resulting estimators are discussed. A Monte Carlo experiment and a forecasting application to an Australian daily electricity generation dataset confirm that the proposed approach significantly improves forecast accuracy with respect to standard forecast combination and reconciliation alternatives.

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Coherent forecast combination for linearly constrained multiple time series".

Jane: Here's the gist of the paper. Linearly constrained multiple time series, encountered in contexts like National Accounts or multilevel frameworks, often require methods to combine multiple incoherent base forecasts into a coherent final result to enhance accuracy and achieve coherence.

Tom: This paper develops an optimization-based technique for coherent forecast combination that combines multiple unbiased base forecasts while respecting the constraints of the time series.

Jane: The authors derive closed-form expressions for both the coherent combined forecast vector (y c = psi y b = M omega y b, where M = I n - W c C C C-1, omega = W-1 K c, and W c = K W-1 K-1) and its error covariance matrix (W fc = M W c M).

Tom: The theoretical foundation relies on the assumptions that base forecasts are unbiased and the error covariance matrix is known, with practical estimation methods like using in-sample residuals outlined.

Paper summary: Tom: So we're looking at this paper now titled "Coherent forecast combination for linearly constrained multiple time series". This whole thing is about taking several separate, sometimes messy forecasts and putting them together in a way that respects the rules of how those variables are linked.

Jane: Exactly. The core idea here is that when you have multiple forecasts for things that are all related by some kind of constraint—like how different parts of an economy fit together—you can get a much better final prediction by combining them coherently instead of just picking the best one or averaging them up blindly.

Lu: What the authors are doing is developing an optimization-based technique to do this combination while making sure the resulting forecast still respects those original constraints, which is what makes it coherent. They aim to combine multiple unbiased base forecasts without violating those time series rules.

Meng: From my side, I wonder how robust this method is when we're talking about real-world data. They claim they've derived closed-form expressions for both the combined forecast vector and its error covariance matrix, which sounds mathematically neat on paper but I need to know what that means for deployment.

Lalam: The paper presents closed form expressions for the coherent combined forecast vector, which is written as y sub c equals Psi transpose y sub b equals M omega transpose y sub b, where M is a matrix defined as one minus W c C C transpose C inverse, and omega is defined as W inverse K c, and W c is equal to K transpose W inverse K inverse >

Tom: That sounds like some heavy linear algebra. So what's the big picture here for us? Why should someone listening on their drive care about this kind of mathematical trick with time series?

Jane: It matters because many real-world problems, like national accounts or energy consumption data, are structured hierarchically. If you only look at one part, it's incomplete. This paper suggests that using all the available forecasts together can improve the accuracy and make the final result more sensible for decision-making.

Lu: The theoretical foundation of this work rests on two main assumptions: first, that all your base forecasts are unbiased, meaning they aren't systematically over- or under-predicting anything, and second, that you already know the error covariance matrix for those base forecasts.

Meng: Knowing the error covariance matrix is a big assumption in practice. How do you get those matrices reliably from real traffic data? What if the structure of that error isn't exactly what the model assumes?

Lalam: The paper suggests using in-sample residuals as a practical way to estimate that error covariance matrix, which is how they make their theoretical framework usable in a setting where we don't know everything beforehand >

Paper summary: Tom: Okay, so they have this theory and some practical estimation methods. What did they actually test? Did they just build this thing and say it works?

Jane: They did empirical validation on the daily Australian electricity generation hierarchical time series. They compared their coherent combination approach against several other things, like individual base forecasts, single-task combinations, and even sequential coherent combinations.

Lu: The results show that the coherent multi-task forecast combination consistently outperformed all those benchmark methods across different settings for both upper and bottom time series >

Meng: Outperformed how many? I need to know the actual performance gain. Are we talking about a small improvement in mean square error, or something substantial enough to justify the complexity of this optimization technique?

Tom: The paper specifically points out that the optimal combination forecast vector, which minimizes the mean square error, is given by y equals Psi transpose y b equals M omega transpose y b >

Jane: That specific combination formula is what they found to be best for minimizing error when you're looking for an unbiased estimator within all possible linear combinations of your base forecasts >

Lu: The study looked at various performance metrics, including the Average Relative Mean Square Error and the Average Relative Mean Absolute Error >

Meng: I see those tables. They compare different forecast horizons and panel sizes. What are some of the caveats they mention about these comparisons? Is this method only good for one specific structure of constraints?

Lalam: They show results across different forecast horizons, panel sizes, and aggregation levels for both upper and bottom time series >

Tom: It seems like the coherent approach, which they call occbe in their simulations on the Australian data, beats methods like ew which is a common benchmark >

Jane: The conclusion of this work is that the problem boils down to finding the best unbiased estimator you can get by linearly combining all your base forecasts, and their method, occbe, came out as the most effective overall technique they tested >

Lu: They are also flagging that future work will focus on more advanced covariance matrix estimation and extending this framework into probabilistic forecasting scenarios >

Meng: So for someone trying to deploy this in production right now, what's the immediate practical hurdle? Is it the computation cost of solving that optimization problem, or is it just trusting the closed-form expressions they gave us?

Tom: The paper doesn't really go into heavy computation cost analysis, but if you have a huge number of variables, solving that optimization might become slow. That's something we'll need to watch >

Jane: It sounds like this is a solid theoretical step forward for combining forecasts in complex systems, moving beyond just picking one model >

Lu: It moves the conversation from simply comparing models to finding the mathematically optimal way to combine multiple uncertain inputs while respecting structural constraints >

Conclusion: Tom: So this paper is about taking several different forecasts for related things—like parts of an economy—and figuring out the best way to stitch them together so the final result makes sense given all those rules.

Jane: It’s titled "Coherent forecast combination for linearly constrained multiple time series," and it’s from some researchers looking at how to combine data when you have these tight structural links between variables.

Lu: The authors are using an optimization method to find that best stitch, aiming to combine forecasts without breaking those original time series constraints.

Meng: From my side, the big question is whether this combination actually holds up when we try to use it for real-time predictions where we can't just look at perfect historical data.

Lalam: I’m looking at the literature and I’m pretty sure the core idea here is that finding this optimal combination is about finding the best unbiased estimator you can get from all your separate base forecasts.

Tom: Finding that best unbiased estimator sounds like it could be really useful if we're dealing with complex systems where a single forecast just isn't enough.

Jane: Exactly, because instead of trusting just one model, this method suggests we can use the information from every available base forecast to get a more accurate picture.

Lu: The paper shows that their method, which they call occbe in their tests on Australian electricity data, actually performs better than just using individual forecasts or combining them sequentially.

Meng: Better how much? Is this just a tiny bit of improvement over what we’re currently doing with simple averaging?

Tom: The authors show it consistently outperforms benchmarks like ew across different time horizons and panel sizes in their simulations on that electricity data set.

Jane: It suggests that for complex, constrained systems, the way you combine the inputs matters just as much as the individual inputs themselves.

Lu: So they are pointing toward finding this optimal linear combination formula to handle those structural dependencies correctly.

Meng: It sounds like a solid theoretical step forward, but I still need to see how computationally expensive solving that optimization actually is when we scale it up to our massive production models.

More episodes

← Home