Chapter 26 Multivariate Forecasting

Chapter 25 developed the Nonlinear Nonparametric Statistics (NNS) approach to univariate forecasting. A single series is partitioned into phase-specific component series, each regressed against its occurrence index and extrapolated using linear or directional nonparametric fits. The multivariate extension retains those univariate forecasts as its first stage.

But many forecasting problems are not univariate.

Macroeconomic indicators move together.
Financial variables transmit shocks across markets.
Operational systems contain multiple indicators observed at different frequencies.
In all such cases, the future of one variable depends not only on its own past, but also on the lagged history of other variables.

This chapter extends the forecasting framework from one time series to a system of time series.

The package implementation is NNS.VAR, a nonparametric vector autoregressive model incorporating NNS.ARMA estimates of variables into NNS.reg for a multivariate time-series forecast. Its purpose is explicit: combine univariate nonlinear time-series forecasting with multivariate nonlinear regression so that each series can borrow information from the others while still retaining its own temporal structure.

Whenever this chapter appeals to mean-split regression behavior (local refinement with occupancy control), the theoretical reference point is Chapter 18: the consistency conditions for recursive mean-split estimation under shrinking local diameters and growing local sample support. Chapter 22 explains the implemented multivariate regression-point matrix and prediction rule, while Chapter 24 develops the ensemble methods employed here.

A second implementation, NNS.nowcast, wraps this framework for mixed-frequency macroeconomic forecasting and nowcasting, using a built-in panel of Federal Reserve style indicators and passing them to NNS.VAR with monthly alignment and a 12-lag structure.

The key ideas of the chapter are therefore:

  • multivariate time-series dependence,
  • nonlinear vector autoregressive extensions,
  • directional cross-variable interactions,
  • mixed-frequency inputs,
  • and empirical forecasting systems built from these components.

26.1 From Univariate to Multivariate Forecasting

Let

\[ X_t = \begin{pmatrix} X_{1t}\\ X_{2t}\\ \vdots\\ X_{pt} \end{pmatrix} \in \mathbb{R}^p \]

denote a \(p\)-dimensional time series.

In univariate forecasting, we model a future value as

\[ X_{t+h} = f(X_t, X_{t-1}, X_{t-2}, \dots) + \varepsilon_{t+h}. \]

In the multivariate setting, each component may depend on the lagged history of every variable in the system:

\[ X_{j,t+h} = f_j\!\bigl( X_{1,t}, X_{1,t-1}, \dots, X_{2,t}, X_{2,t-1}, \dots, X_{p,t}, X_{p,t-1}, \dots \bigr) + \varepsilon_{j,t+h}, \qquad j=1,\dots,p. \]

So the problem is no longer to forecast a single path in isolation. It is to estimate a joint dynamic structure in which variables affect one another through time.

This is the motivation for vector autoregression.


26.2 Classical VAR and Its Limits

A classical vector autoregressive model of order \(k\), VAR\((k)\), is written

\[ X_t = c + A_1 X_{t-1} + A_2 X_{t-2} + \cdots + A_k X_{t-k} + \varepsilon_t, \]

where \(c\) is a vector of intercepts, \(A_1,\dots,A_k\) are coefficient matrices, and \(\varepsilon_t\) is an innovation vector.

This formulation is powerful because it allows each variable to depend on lagged values of all the others. But it also imposes several strong restrictions.

26.2.1 Linearity

All effects enter linearly through the coefficient matrices. Threshold behavior, asymmetry, nonlinear interactions, and regime changes are not modeled directly.

26.2.2 Uniform lag structure

Classical VAR usually imposes the same lag order across all variables, even when different variables operate naturally at different horizons.

26.2.3 Parametric residual thinking

Forecast construction and inference are built around residual assumptions that may become fragile in heavy-tailed, asymmetric, or nonlinear environments.

26.2.4 Mixed-frequency difficulty

Variables observed monthly, quarterly, weekly, or daily do not fit naturally into the same linear lag system without additional modeling layers.

The NNS approach keeps the useful intuition of cross-variable lag forecasting while relaxing these structural restrictions.


26.3 Nonparametric Vector Autoregression

NNS.VAR reframes vector autoregression as a nonlinear nonparametric learning problem.

The implementation has a clear architecture.

26.3.1 Stage 1: complete each series individually

Each variable’s missing values are first estimated from its own observations, and each completed series is forecast forward with NNS.ARMA. When a variable’s latest values are missing, every missing value comes from an NNS.stack fit on the time index: it takes the value of the nearest regression point on the index, the newest observation keeps its own regression point, and values after it extend the gradient at that boundary. When only earlier values are missing, they are interpolated linearly. The returned object explicitly includes

  • interpolated_and_extrapolated,
  • univariate,
  • multivariate,
  • ensemble,
  • and relevant_variables.

This design is important. The multivariate forecast does not replace the univariate forecast. It is built on top of it.

26.3.2 Stage 2: construct lagged predictors

The completed panel is extended with the univariate forecasts and transformed into a lag matrix. At each forecast date, unknown same-period covariates and unknown lagged values come from these first-stage forecasts. The target variable’s own contemporaneous value is excluded; other series’ contemporaneous predictions remain part of the intended two-stage design.

26.3.3 Stage 3: reduce lagged predictors

For each target variable, the lagged predictor set is screened using one of several relevance measures and a thresholding rule.

26.3.4 Stage 4: estimate nonlinear multivariate forecasts

The retained lagged predictors are passed into the multivariate regression system through NNS.stack and NNS.reg.

26.3.5 Stage 5: combine univariate and multivariate information

The final forecast is an ensemble of the univariate and multivariate components.

So the “VAR” terminology remains conceptually appropriate, but the mechanism is no longer a linear matrix recursion. It is a nonlinear regression surface over lagged multivariate data.

The Python workflow mirrors the five-stage architecture through nns_var, which returns the same five-part structure:

import numpy as np
from nns import nns_var

rng = np.random.default_rng(123)
n = 200
x1 = np.cumsum(rng.normal(size=n))
lag3_x1 = np.concatenate((np.repeat(x1[0], 3), x1[:-3]))
x2 = 0.5 * lag3_x1 + rng.normal(size=n)
variables = np.column_stack((x1, x2))

out = nns_var(variables, h=12, tau=12, naive_weights=True, status=False)

out["interpolated_and_extrapolated"]   # completed panel
out["relevant_variables"]              # selected lagged predictors per target
out["univariate"]                      # standalone NNS.ARMA forecasts
out["multivariate"]                    # nonlinear multivariate forecasts
out["ensemble"]                        # combined forecast

# For an already aligned mixed-frequency panel:
# out = nns_var(econ_variables, h=12, tau=12, nowcast=True)
# The Python flag is for API compatibility; it does not download the panel.

26.4 Lagged Predictor Geometry

Suppose the observed system is stored in a matrix

\[ V = \begin{bmatrix} X_{11} & X_{21} & \cdots & X_{p1}\\ X_{12} & X_{22} & \cdots & X_{p2}\\ \vdots & \vdots & \ddots & \vdots\\ X_{1T} & X_{2T} & \cdots & X_{pT} \end{bmatrix}. \]

The first structural step is to generate lagged copies of each column. For a lag depth \(\tau\), these predictors include terms such as

\[ X_{j,t},\; X_{j,t-1},\; X_{j,t-2},\; \dots,\; X_{j,t-\tau}. \]

A major advantage of NNS.VAR is that the lag argument tau is flexible.

It may be

  • a single positive integer, applying the same lag depth to every variable,
  • a vector, assigning a single lag choice to each variable,
  • or a list, assigning multiple lags to each variable separately.

Thus the lag structure need not be homogeneous.

One variable may use only lag 1, another may use lags 1 through 6, and a third may use a sparse pattern such as \(1, 3, 12\). This flexibility is especially important in economic and financial systems, where different variables evolve over different time scales.

At forecast origin \(T\), define the extended series

\[ \widetilde X_{k,s\mid T}= \begin{cases} X^{\mathrm{completed}}_{k,s\mid T}, & s\le T,\\ \widehat u_{k,s\mid T}, & s>T. \end{cases} \]

For target \(j\) and horizon \(h\), the feature vector \(\widetilde Z_{j,T+h\mid T}\) contains the selected entries \(\widetilde X_{k,T+h-\ell\mid T}\), excluding \((k,\ell)=(j,0)\). The multivariate forecast is

\[ \widehat m_{j,T+h\mid T}=f_j(\widetilde Z_{j,T+h\mid T}). \]

The surface \(f_j\) is fitted to historical rows with target \(X_{j,t}\) and the corresponding contemporaneous cross-series and lagged predictors. Deployment then substitutes first-stage predictions where those inputs are unknown. This is a plug-in forecasting architecture, not a separate direct-horizon regression \(X_{j,T+h}=f_j(Z_T)\). At later horizons, even positive-lag features can lie beyond \(T\) and must be forecast as well.


26.5 Mixed-Frequency Inputs

A central practical challenge in multivariate forecasting is mixed-frequency data.

Examples include:

  • monthly inflation with quarterly GDP,
  • weekly claims with monthly industrial production,
  • daily market variables with monthly macroeconomic releases.

If these series are aligned onto a common calendar, the lower-frequency variables necessarily contain missing entries at higher-frequency timestamps.

Classical multivariate methods often require specialized machinery for this. The NNS framework instead treats the problem as a nonlinear interpolation and extrapolation task before multivariate estimation begins.

Let

\[ V_t = (X_{1t},\dots,X_{pt}), \]

with some entries unobserved because the reporting frequencies differ. NNS.VAR explicitly returns a matrix called interpolated_and_extrapolated, whose purpose is to replace missing values in the original panel with a nonparametric fit of each variable on the time index. This is not an incidental preprocessing step. It is foundational to the mixed-frequency design of the method.

Conceptually, the procedure is:

  1. place all series on a common index,
  2. complete each series from its own observations: when its latest values are missing, with an NNS.stack fit on the index that also carries the series to the end of the panel, and otherwise by linear interpolation,
  3. build the lagged multivariate system on the completed panel.

So mixed-frequency forecasting is handled natively, rather than being outsourced to a separate parametric state-space model.


26.6 Dimension Reduction and Directional Relevance

Once the lagged panel has been created, not all lagged predictors are equally informative for each target variable.

NNS.VAR therefore applies a dimension-reduction step before forming the multivariate forecast. The user can choose among four relevance schemes:

  • "cor": signed Spearman rank-correlation coefficients, screened by magnitude,
  • "NNS.dep": nonlinear dependence weights,
  • "NNS.caus": directional association weights,
  • "all": the average of the Spearman, dependence, directional-association, and unit-weight coefficients.

This is a substantive departure from classical VAR, where the usual practice is to include all lags up to the chosen order unless restrictions are imposed manually.

Here the guiding question is:

Which lagged variables are actually relevant for forecasting this target?

The implementation answers this in two steps.

26.6.1 Step 1: compute relevance scores

For a target variable \(Y\) and lagged predictors \(Z_1,\dots,Z_m\), a coefficient \(w_\ell\) is computed for each predictor using the selected method. Screening uses its magnitude \(r_\ell=|w_\ell|\), while the synthetic index retains its sign.

For dim.red.method = "cor", the screening score is

\[ r_\ell = |\rho_S(Y,Z_\ell)|, \]

For the other methods, the score is the magnitude of the nonlinear dependence coefficient, the directional association coefficient, or the four-component average described above.

26.6.2 Step 2: tune and apply a threshold before fitting Stack

The relevance helper fits NNS.reg(type = "REG") at each candidate coefficient cutoff and scores its validation forecasts using obj.fn. It selects the cutoff with the best objective. Under the default validation = "pipeline", the candidate grid and fitted coefficients come from the training prefix, and the validation predictors are generated from that same forecast origin. NNS.stack is called only after this screening has selected the predictor columns.

Outside the fallback case below, the selection rule is

\[ Z_\ell \text{ is retained} \quad \Longleftrightarrow \quad r_\ell > 0 \quad\text{and}\quad r_\ell \ge \theta, \]

where \(\theta\) is the cutoff selected by the direct NNS.reg validation search. Equality with a positive cutoff retains the coefficient; zero coefficients do not enter the selected index.

If thresholding removes every coefficient, the regression restores the unthresholded coefficients. If those are all zero, unit coefficients retain the full predictor set. The selected columns are then passed to NNS.stack(method = 1, stack = FALSE) for multivariate tuning and forecasting.

This deserves emphasis. The reduction step is not just “screening by method name.” It is a concrete thresholding rule applied to lag-specific relevance scores.


26.7 Directional Cross-Variable Interactions

Why are these relevance measures important?

Because cross-variable forecasting effects are not generally linear, symmetric, or monotone.

A lagged predictor may matter because:

  • it moves with the target linearly,
  • it influences the target nonlinearly,
  • or it exhibits directional causal strength that would be poorly summarized by a simple coefficient.

The NNS framework is designed precisely for such cases.

If variable \(X_k\) influences variable \(X_j\) only during downturns, only above a threshold, or only through asymmetric responses, then a linear VAR coefficient can understate or even obscure the relationship. By contrast, dependence weights from NNS.dep and directional weights from NNS.caus are built to recognize these nonlinear and asymmetric structures.

So the multivariate forecasting problem becomes one of discovering directional cross-variable interactions in lag space.


26.8 Multivariate Forecasting as Nonlinear Regression

For each target variable \(X_j\), the forecasting problem can be written as

\[ \widehat m_{j,T+h\mid T}=f_j(\widetilde Z_{j,T+h\mid T}), \]

where \(\widetilde Z_{j,T+h\mid T}\) is the selected extended-panel feature vector after dimension reduction.

The surface \(f_j\) is estimated through the NNS regression framework rather than through a linear coefficient matrix. This is the conceptual heart of NNS.VAR.

Multivariate autoregression is therefore reinterpreted as a supervised learning problem:

  • the training response is a historical value of one variable,
  • the predictors include other series at that date and the selected lags,
  • future rows replace unavailable predictors with first-stage forecasts,
  • the regression surface is estimated nonparametrically.

This is still autoregression, but with nonlinear function estimation replacing linear matrix multiplication.


26.9 Ensemble Construction

One of the most important features of NNS.VAR is that the final output is not merely a multivariate forecast.

For each target variable, the method returns:

  • a univariate forecast from NNS.ARMA,
  • a multivariate forecast from the nonlinear regression stack,
  • and an ensemble combining the two.

Let

\[ \hat u_{j,t+h} \]

denote the univariate forecast for variable \(j\), and let

\[ \hat m_{j,t+h} \]

denote the multivariate forecast. The ensemble is

\[ \hat x_{j,t+h} = w^{(u)}_j \hat u_{j,t+h} + w^{(m)}_j \hat m_{j,t+h}, \qquad w^{(u)}_j + w^{(m)}_j = 1. \]

If naive.weights = TRUE, the method uses equal weights:

\[ w^{(u)}_j = w^{(m)}_j = \frac12. \]

If naive.weights = FALSE, the implementation is more precise than a simple “count of relevant variables.” It computes the proportion of selected lagged predictors that belong to the target variable itself.

Let

  • \(n_{\text{own},j}\) be the number of retained lagged predictors for target \(j\) whose base variable is \(j\),
  • \(n_{\text{cross},j}\) be the number of retained lagged predictors whose base variable is not \(j\).

Then

\[ w^{(u)}_j = \frac{n_{\text{own},j}}{n_{\text{own},j} + n_{\text{cross},j}}, \qquad w^{(m)}_j = 1 - w^{(u)}_j = \frac{n_{\text{cross},j}}{n_{\text{own},j} + n_{\text{cross},j}}. \]

This is an elegant weighting rule.

  • If the selected structure is mostly own-lags, the forecast leans univariate.
  • If the selected structure is mostly cross-variable, the forecast leans multivariate.
  • If there is no usable distinction, the implementation falls back to equal weighting.

So the ensemble is not arbitrary averaging. It is a structural weighting mechanism derived from the selected lag geometry.


26.10 Relevant Variables as a Structural Map

The output relevant_variables provides more than a convenience list. It gives a structural summary of the system learned by the model.

For each target variable, it records the lagged predictors that survived the relevance threshold. This means the forecast comes with an interpretable dynamic footprint:

  • which own-lags matter,
  • which other variables matter,
  • and at which lag horizons they matter.

This is a predictive relevance map. Because the model includes contemporaneous cross-series inputs and selects features for forecasting, the map alone is not a Granger-causality test or a causal identification result.

A variable may be forecast mainly by its own lagged history, by a sparse collection of related series, or by a broad network of interacting indicators. The model does not impose one of these patterns a priori; it discovers them from the data.


26.11 A Small Synthetic Illustration

Suppose we jointly forecast monthly inflation \(I_t\), unemployment \(U_t\), and industrial production \(P_t\), using lags through month 12.

The full lag system would contain predictors such as

\[ I_{t-1}, I_{t-2}, \dots, I_{t-12},\; U_{t-1}, U_{t-2}, \dots, U_{t-12},\; P_{t-1}, P_{t-2}, \dots, P_{t-12}. \]

But after dimension reduction, the retained structures for different targets need not look alike.

For inflation, the model might retain mainly

\[ I_{t-1},\ I_{t-2},\ P_{t-1}, \]

indicating that short-run inflation persistence and recent production conditions are the dominant signals.

For unemployment, the model might retain

\[ U_{t-1},\ U_{t-12},\ I_{t-3}, \]

suggesting a mixture of local persistence, annual seasonality, and delayed inflation spillover.

This synthetic example illustrates the main point: the NNS system does not treat all targets as sharing the same lag law. Each response learns its own sparse multivariate structure.

That is exactly what a nonlinear vector autoregression should do.


26.12 Nowcasting with NNS.VAR

Nowcasting deserves special treatment because it is one of the most practical uses of the framework.

NNS.nowcast is a wrapper built around NNS.VAR. It downloads a base panel of macroeconomic indicators, converts each to monthly frequency, merges them into a common panel, and then calls NNS.VAR(econ_variables, h = h, tau = 12, nowcast = TRUE, naive.weights = naive.weights).

This gives a ready-made mixed-frequency multivariate forecasting system.

The built-in variable set includes indicators such as:

  • payroll employment,
  • job openings,
  • CPI and core CPI,
  • durable goods orders,
  • retail sales,
  • unemployment,
  • housing starts and permits,
  • industrial production,
  • personal income,
  • exports and imports,
  • construction spending,
  • unit labor cost,
  • real consumption spending,
  • real GDP,
  • weekly unemployment claims,
  • Treasury rates and yield spreads,
  • Federal Reserve balance sheet assets,
  • commodity prices,
  • federal funds,
  • producer prices,
  • labor force participation,
  • money supply,
  • and ADP payrolls.

This is a practically meaningful nowcasting panel. It combines labor, inflation, output, rates, liquidity, spending, trade, and commodity signals in one monthly-aligned system.

26.12.1 What the nowcast output means

The output of NNS.nowcast is not a single number. It returns the five principal forecast components from NNS.VAR, together with validation metadata:

  • interpolated_and_extrapolated: the completed monthly panel after filling mixed-frequency gaps,
  • relevant_variables: the selected lagged predictors for each target,
  • univariate: the standalone NNS.ARMA forecasts,
  • multivariate: the nonlinear multivariate forecasts,
  • ensemble: the combined forecast.

So in practice, nowcasting means more than “guess the current GDP number.” It means:

  1. align all indicators to the current monthly calendar,
  2. fill missing lower-frequency observations,
  3. estimate joint nonlinear lag relations,
  4. produce univariate, multivariate, and ensemble projections,
  5. inspect which variables were actually driving the current nowcast.

This is a much richer object than a single real-time estimate.


26.13 Prediction Intervals

As in Chapter 17, point forecasts alone are not enough. Forecast uncertainty must also be quantified.

The examples associated with both NNS.VAR and NNS.nowcast construct prediction intervals using NNS.meboot, LPM.VaR, and UPM.VaR.

Suppose the forecasted path of a target series is

\[ \hat x_1,\hat x_2,\dots,\hat x_h. \]

Bootstrap replicates are generated from the forecast path:

\[ \hat x^{*(1)}, \hat x^{*(2)}, \dots, \hat x^{*(B)}. \]

Lower and upper forecast bounds are then obtained using partial-moment quantile operators:

\[ \text{Lower}_{\alpha} = \operatorname{LPM.VaR}(\alpha/2,\cdot), \qquad \text{Upper}_{\alpha} = \operatorname{UPM.VaR}(\alpha/2,\cdot). \]

In the nowcasting examples, this is illustrated directly for GDP. The GDP ensemble path is bootstrapped with NNS.meboot, and lower and upper confidence bounds are then computed from LPM.VaR and UPM.VaR.

Because the replicates are generated from the forecast path, these bounds describe the dispersion of the path rather than the uncertainty of each future outcome: they do not widen with the horizon and are not calibrated to a coverage probability (see Chapter 25).

The conceptual advantage is consistent with the rest of NNS:

  • no Gaussian residual assumption is required,
  • asymmetry is naturally allowed,
  • and dependence-preserving synthetic replicates can be generated rather than relying on iid residual resampling.

26.14 Comparison with Classical VAR

The difference between classical VAR and NNS.VAR can be summarized clearly.

Feature Classical VAR NNS.VAR
Functional form Linear Nonlinear nonparametric
Lag structure Usually common across variables Scalar, vector, or list-based
Missing mixed-frequency values Requires extra modeling machinery Built into interpolation and extrapolation
Variable selection Often manual or penalized Built-in relevance screening with thresholding
Multivariate dependence Linear coefficients Dependence and causation-aware feature selection
Forecast output One system forecast Univariate, multivariate, relevant-variable map, and ensemble
Ensemble weighting Usually absent Equal or structure-based own-lag versus cross-lag weighting
Prediction intervals Parametric or residual bootstrap Partial-moment quantile intervals with NNS.meboot

The NNS version therefore preserves the idea that variables should be modeled jointly, but relaxes the linear and parametric restrictions around that idea.


26.15 Why the Directional Framework Matters

The contribution of the directional framework is not merely technical. It changes what multivariate forecasting can represent.

26.15.1 Nonlinearity is primary

Forecast relationships do not need to be approximated by a global linear law.

26.15.2 Cross-variable effects can be asymmetric

Dependence and causation can be directional, state-dependent, and nonlinear.

26.15.3 Mixed-frequency panels become tractable

Incomplete higher-frequency panels are treated as a forecasting geometry problem rather than as a separate special case.

26.15.4 Univariate structure is preserved

Each variable retains its own autoregressive signature through the univariate component.

26.15.5 Forecast combination is structural

The ensemble reflects how much of the retained lag structure is own-history versus cross-variable history.

These are not minor refinements. Together they turn vector autoregression into a flexible directional learning architecture.


26.16 Empirical Applications

The framework is naturally suited to systems in which variables move together, asymmetrically, and at different reporting frequencies.

26.16.1 Macroeconomics

GDP, employment, inflation, industrial production, claims, rates, and spending variables are all natural candidates for mixed-frequency nonlinear nowcasting.

26.16.2 Finance

Asset returns, volatility, spreads, rates, and macro variables often interact through threshold effects and asymmetric transmission.

26.16.3 Operations and supply chains

Demand, inventories, labor, orders, and shipping activity may be observed on different calendars but still require joint forecasting.

26.16.4 Energy and commodities

Production, inventories, spot prices, futures curves, and macro demand indicators evolve jointly but not linearly.

In all these settings, the problem is not merely to fit a system. It is to discover how a dynamic system actually transmits information through time.


26.17 Leakage-Safe Validation in Multivariate and Mixed-Frequency Settings

Before evaluating forecast accuracy, the data pipeline itself must be audited for information timing and release-order integrity.

NNS.VAR(validation = "pipeline") uses one chronological tuning origin inside the supplied history. It completes each raw prefix independently, forecasts every series across the heldout window, and builds both contemporaneous and lagged heldout features from that extended prefix. Relevance coefficients and candidate grids are fitted on the prefix, and only actually observed heldout targets are scored; imputed target values do not count as validation outcomes. The final selected regression is then refitted on all available observations.

The returned validation entry states the origin, horizon and observed target counts. validation = "observed" retains the earlier tuning calculation with realized holdout covariates for reproducibility. Its score measures conditional regression performance given those covariates, rather than the deployed two-stage forecast.

A single internal tuning origin is not an independent performance report. To evaluate accuracy, repeat the complete pipeline at multiple outer rolling origins, including completion, univariate model selection, dimension reduction, tuning and final forecast combination. This also evaluates errors in the first-stage covariates.

In multivariate nowcasting/forecasting, leakage control is more delicate because indicators arrive asynchronously.

Use the following rules:

  • Vintage-consistent features: at forecast origin t, include only indicator values actually released by t.
  • Release-calendar alignment: construct mixed-frequency regressors using publication timestamps, not finalized revised datasets.
  • Origin-wise completion and forecasting: fill gaps from the available prefix and forecast every unknown feature, including later-horizon lags.
  • Origin-wise dimension reduction: relevance screening/thresholding must be recomputed within each training window.
  • Horizon-by-horizon evaluation: report forecast and interval metrics by target horizon and by target variable.
  • Stability diagnostics: track how the relevant_variables set changes over origins to distinguish signal from selection noise.

These rules make NNS.VAR/NNS.nowcast evaluations comparable to production conditions and prevent optimistic bias from look-ahead information.


26.18 Summary

This chapter extended the NNS forecasting framework from univariate series to multivariate forecasting.

The main results are:

  • A multivariate time series is treated as a nonlinear autoregression problem in lagged multivariate space.
  • NNS.VAR combines univariate NNS.ARMA forecasts with multivariate nonlinear regression forecasts.
  • The lag structure is flexible: tau may be a scalar, vector, or list.
  • Mixed-frequency inputs are handled through interpolation and extrapolation before constructing the lagged system.
  • Dimension reduction may be based on correlation, nonlinear dependence, directional causation, or their average.
  • Relevance cutoffs are selected by a direct NNS.reg validation search; the surviving predictor columns are then passed to NNS.stack for multivariate tuning and forecasting.
  • The ensemble forecast is either equally weighted or weighted according to the share of retained predictors that are own-lags versus cross-variable lags.
  • In classification-style forecasting tasks with skewed targets, NNS ensemble workflows can combine minority up-sampling and majority down-sampling across ensemble members to reduce imbalance bias.
  • The output relevant_variables provides a structural map of the learned dynamic system.
  • NNS.nowcast applies this architecture directly to a practical macroeconomic panel with mixed frequencies and monthly alignment.
  • NNS.meboot, LPM.VaR and UPM.VaR describe forecast-path scenario bands; outcome prediction intervals additionally require calibration.

Classical vector autoregression taught an important lesson: forecasting improves when variables are modeled jointly. The NNS framework retains that insight but frees it from linearity, common-frequency restrictions, and rigid parametric structure.

The result is a forecasting system for nonlinear, asymmetric, mixed-frequency multivariate data:

nonparametric vector autoregression as directional multivariate learning through time.