Nonlinear Probabilistic Forecast Reconciliation Journal: International Journal of Forecasting
Anubhab Biswas, Lorenzo Zambon, Lorenzo Nespoli, Giorgio Corani · 2026
Read Nonlinear Probabilistic Forecast Reconciliation Journal: International Journal of Forecasting by Anubhab Biswas, Lorenzo Zambon, Lorenzo Nespoli, Giorgio Corani (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-26668.txt.
Introduction
Biswas, Zambon, Nespoli, and Corani tackle forecast reconciliation under nonlinear constraints, a case previous methods did not cover. They extend both projection, which maps forecast samples onto the coherent manifold, and conditioning, which constrains the joint forecast. The conditioning approach is built with an Unscented Kalman Filter method. On made-up and real data sets, both methods improve accuracy in general, with the filter-based approach best overall, fastest, and always kept among the plausible models.
Catalog record
| CREATOR | Anubhab Biswas, Lorenzo Zambon, Lorenzo Nespoli, Giorgio Corani |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 44904 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.26668
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Nonlinear Probabilistic Forecast Reconciliation Journal: International Journal of Forecasting Anubhab Biswas, Lorenzo Zambon, Lorenzo Nespoli, Giorgio Corani Abstract Forecast reconciliation adjusts independently generated forecasts so that they satisfy some known constraints. While probabilistic forecast reconciliation is well established for linear constraints, some practical forecasting problems involve nonlinear relationships among variables. In this paper, we address probabilistic forecast reconciliation with nonlinear constraints for the first time. We extend both reconciliation via projection and conditioning to the case of nonlinear constraints. The projection approach reconciles forecast samples by mapping them onto the nonlinear coherent manifold. The conditioning approach adopts a sampling algorithm inspired to the Unscented Kalman Filter (UKF). We evaluate both methods on synthetic and real datasets. Empirically, both reconciliation approaches generally improve forecast accuracy. The UKF-based approach achieves the best overall performance while being substantially faster than the projection one. 1 Introduction Time series are often related by linear constraints of aggregation. For example, for a certain store, consider forecasting both the sales of each item and the total sales. The forecast of the total sales of the store should be equal to the sum of the forecast of the sales of each item. If forecasts satisfy these constraints they are coherent , while if they fail to satisfy the constraints they are incoherent . Base forecasts , which are created independently for each time series, are generally incoherent. Forecast reconciliation adjusts the base forecasts making them coherent. The first methods for forecast reconciliation were based on projection ( 1 ; 9 ) . In particular, 20 obtained in closed form the projection that minimizes the expected squared error of the coherent forecast. Such methods can be interpreted as projecting the base point forecast onto the linear coherent subspace ( 16 ) . While the above methods refer to point forecast reconciliation, the most advanced approaches perform linear probabilistic reconciliation yielding a probability density on the linear coherent subspace. Earlier studies include 18 , where an empirical copula based approach for reconciliation which uses permutation of the samples was studied. There are two main approaches to probabilistic forecast reconciliation. The first one is the extension of the reconciliation via projection method to the probabilistic case ( 17 ; 21 ) . The second approach is reconciliation via conditioning where the incoherent density of the base forecast is sliced on the coherent subspace. This approach has been extensively studied for continuous, discrete and mixed-type of linearly constrained time series ( 23 ; 22 ; 24 ) . Also the probabilistic approaches of 3 ; 4 have been shown to be equivalent to reconciliation via conditioning ( 22 ) . Several applications in multiple time series forecasting include nonlinear constraints between the time series, such as socio-demographic ratios like mortality or unemployment rates ( 6 ) . Very recently the projection-based reconciliation approach has been extended to obtain coherent forecasts for time series with nonlinear constraints ( 14 ; 6 ) ; however, these works are limited to point forecasts. We address probabilistic forecast reconciliation under nonlinear constraints for the first time. We extend both reconciliation via projection and via conditioning to the case of time series with nonlinear constraints. The projection-based reconciliation projects each sample point from the cloud of the base forecast distribution to points on the manifold. However, it has the shortcoming of requiring a numerical optimization for each sample to be reconciled, leading to heavy computation times. yoriichi300As an alternative, we also propose an algorithm based on reconciliation via conditioning , with a sampling strategy inspired by the Unscented Kalman Filter ( 12 ) . In our experiments, it provides slightly more accurate results than projection while also substantially reducing computation time. The paper is organized as follows. Section 2 introduces the notations and terminology needed for probabilistic nonlinear reconciliation. Section 3 introduces probabilistic reconciliation under nonlinear constraints and presents the extensions of reconciliation via projection and reconciliation via conditioning. Section 4 discusses the experimental setup and evaluates the proposed methods on both simulated synthetic data with different nonlinearities and two real case studies involving socio-demographic ratios of Switzerland and tourism ratio of Australia. Section 5 concludes and highlights the directions for future research along with the limitations of our proposed methods. All the codes related to the experiments can be found in the public Github repository 1 1 1 https://github.com/ijfpnlr-svg/ijf_sub and can be reproduced. The reported computation times were obtained on a machine with Intel Core i7-1370P processor (20 CPU cores) and 32 GB of RAM, running in Linux . 2 Notation Z=\frac{X}{Y} X Y Figure 1 : Example of time series with nonlinear constraints; the free time series are X and Y , the constrained time series Z is the ratio of the two. Let us consider the time series \mathbf{B}_{t}=\{B^{1}_{t},\dots,B^{n_{b}}_{t}\}\in\mathbb{R}^{n_{b}} and \mathbf{U}_{t}=\{U^{1}_{t},\dots,U^{n_{u}}_{t}\}\in\mathbb{R}^{n_{u}} , such that, at each time t=1,\dots,T : \mathbf{U}_{t}=f_{u}(\mathbf{B}_{t}), (1) for some function f_{u}:\mathbb{R}^{n_{b}}\to\mathbb{R}^{n_{u}} . This formulation generalizes the classic framework of hierarchical time series, in which f_{u}(\mathbf{B}_{t})=\mathbf{A}\mathbf{B}_{t} and \mathbf{A} is the aggregation matrix. We refer to \mathbf{B}_{t} and \mathbf{U}_{t} respectively as free and constrained time series, and we call f_{u} the free-to-constrained (FTC) function. We show an example in Fig. 1 , where f_{u}(x,y)=x/y . Note that the choice for the set of free time series might be arbitrary; in this case, we could equivalently set X_{t} as the constrained series and f_{u}(y,z)=y\cdot z . We then define the full vector of time series as \mathbf{Y}_{t}=\begin{bmatrix}\mathbf{U}_{t}\\ \mathbf{B}_{t}\end{bmatrix}=\begin{bmatrix}f_{u}(\mathbf{B}_{t})\\ \mathbf{B}_{t}\end{bmatrix}=f(\mathbf{B}_{t}), (2) where f:\mathbb{R}^{n_{b}}\to\mathbb{R}^{n} is the free-to-all (FTA) function. At each time t , the vector \mathbf{Y}_{t} lies on the nonlinear manifold \mathcal{M}:=\{f(\mathbf{b}):\mathbf{b}\in\mathbb{R}^{n_{b}}\}, (3) which generalizes the linear coherent subspace \mathcal{S} used in linear reconciliation. We summarize the notation in Table 1 , where we also compare it with its linear counterpart, keeping similar notations wherever possible. Let \hat{\mathbf{y}}_{t+h} denote the base point forecast of \mathbf{Y}_{t} at horizon h , and \tilde{\mathbf{y}}_{t+h} the reconciled (coherent) point forecast. We denote the base point forecasts for the free and constrained components as \hat{\mathbf{b}}_{t+h} and \hat{\mathbf{u}}_{t+h} , and their reconciled counterparts as \tilde{\mathbf{b}}_{t+h} and \tilde{\mathbf{u}}_{t+h} . In our experiments we focus on h=1 ; and in the following sections we drop the subscript t+h for simplicity. Linear Entity Notation Generalization Notation Bottom time series \mathbf{B}_{t} Free time series \mathbf{B}_{t} Upper time series \mathbf{U}_{t} Constrained time series \mathbf{U}_{t} Coherent linear subspace \mathcal{S} Coherent manifold \mathcal{M} Aggregation matrix \mathbf{A} Free-to-constrained (FTC) function f_{u}(.) Summing matrix \mathbf{S} Free-to-all (FTA) function f(.) No. of bottom series n_{b} No. of free series n_{b} No. of upper series n_{u} No. of constrained series n_{u} No. of series n No. of series n Table 1 : A generalized notation for forecast reconciliation Figure 2 : Example of nonlinear probabilistic reconciliation of the ratio constraint described in Fig. 1 - The purple surface denotes the coherent manifold where the incoherent red points are made coherent as green. Following 17 , we denote the forecast distribution of the free time series B_{t} by the probability triple (\mathbb{R}^{n_{b}},\mathcal{F}_{\mathbb{R}^{n_{b}}},\nu) , where \mathcal{F}_{\mathbb{R}^{n_{b}}} denotes the Borel \sigma -algebra associated with \mathbb{R}^{n_{b}} . Below we define coherence in terms of probability distributions for time series with nonlinear constraints. Definition 1 ( Coherent Probability Triple ) . A probability triple (\mathcal{M},\mathcal{F}_{\mathcal{M}},\breve{\nu}_{\mathcal{M}}) is