Chapter 25 Nonparametric Time Series Models
This chapter turns the directional framework toward time.
Previous chapters developed nonparametric estimation, clustering, regression, classification, and ensemble learning using recursive partitioning, local averaging, and benchmark-relative structure. Those methods treated observations as unordered or cross-sectional. Time-series analysis adds a new constraint:
the observations arrive in sequence, and that ordering matters.
A time series is not merely a set of values. It is a structured sequence in which the past may influence the future, seasonal patterns may recur, and dependence may change across time.
Classical time-series analysis addresses these problems through models such as ARIMA, ETS, and state-space methods. These are often effective, but they inherit familiar limitations:
- linear autoregressive structure,
- parametric error assumptions,
- explicit stationarity requirements,
- and model identification choices that must be imposed before estimation.
The Nonlinear Nonparametric Statistics (NNS) framework approaches time series differently.
At its heart, time-series modeling is treated as a subset regression problem: a sequence is decomposed into lagged component series, and future values are forecast by applying the same nonlinear nonparametric regression logic developed earlier in the book to those components. In this view, autoregression is not abandoned. It is generalized.
Rather than beginning with a linear dynamic equation, NNS begins with a simpler principle:
forecast the future by learning the structure of the past without imposing a parametric law for that structure.
This chapter develops that idea.
25.1 Time Series as Ordered Nonparametric Data
Let
\[ \{X_t\}_{t=1}^T \]
denote a real-valued time series.
In classical analysis, a time series is often modeled through a difference equation such as
\[ X_t = \phi_1 X_{t-1} + \cdots + \phi_p X_{t-p} + \varepsilon_t, \]
possibly after differencing, detrending, or seasonal adjustment.
That formulation assumes from the outset that the dynamic relation is linear in the lagged values.
The directional framework takes a broader view.
A time series can be written as a regression problem in which the response is the future observation and the predictors are functions of the past:
\[ X_t = f(X_{t-1}, X_{t-2}, \dots) + \varepsilon_t. \]
The task is then to estimate the unknown dynamic map \(f\) nonparametrically.
This places time-series analysis inside the same framework developed in Chapters 18–19 and 21–24:
- identify informative local structure,
- partition or decompose the data,
- estimate conditionally,
- and aggregate only afterward.
The only additional element is temporal order.
25.2 Why Classical Time-Series Models Can Fail
The central models of classical forecasting are powerful, but they are built around structural assumptions that many real series violate.
25.2.1 Linearity
ARIMA and related autoregressive models assume that the next observation depends linearly on lagged values, perhaps after transformation. But many series exhibit threshold effects, asymmetric responses, cyclical distortions, or nonlinear seasonal interactions.
25.2.2 Stationarity requirements
The Box–Jenkins framework is built around stationarity. In practice, many observed series are not stationary in level, variance, or seasonal structure. Transformations and differencing may help, but they also alter the object being modeled.
25.2.3 Parametric identification
Classical modeling requires choosing model orders, differencing levels, seasonal terms, and error structures. These decisions can materially change the forecast.
25.2.4 Symmetric error treatment
Least-squares fitting treats positive and negative forecast errors symmetrically, even in contexts where underprediction and overprediction have different consequences.
The directional nonparametric approach seeks to preserve the useful idea of autoregression while relaxing these restrictions.
25.3 Autoregression as a Subset Regression Problem
The NNS time-series framework begins from a simple decomposition.
Suppose a series exhibits a seasonal or cyclical lag \(m\). Then observations separated by that lag belong to a common component series:
\[ \{X_1, X_{1+m}, X_{1+2m}, \dots\}, \quad \{X_2, X_{2+m}, X_{2+2m}, \dots\}, \quad \dots \quad \{X_m, X_{2m}, X_{3m}, \dots\}. \]
Each component series is itself a smaller time series indexed by occurrence number within that phase.
Forecasting then becomes a regression problem on each component series separately.
For a given component series with index vector
\[ z = 1,2,\dots,n_j \]
and values
\[ y^{(j)}_1, y^{(j)}_2, \dots, y^{(j)}_{n_j}, \]
we estimate the next value through either linear or nonlinear regression:
\[ y^{(j)}_{n_j+1} \approx \hat f_j(n_j+1). \]
The final forecast aggregates these component forecasts using weights determined by predictive strength.
This is why time series in NNS are best viewed as a subset regression problem:
- the original series is partitioned into lag-defined subsets,
- each subset is modeled with NNS regression,
- and the subset forecasts are recombined.
Autoregression is therefore retained, but its mechanism is generalized from linear lag equations to nonparametric lag-structure estimation.
A small implementation detail is worth noting. If the total series length \(T\) is not an exact multiple of \(m\), then the component series need not all have the same length. Some phases will contain one more observation than others. This creates no conceptual difficulty, but it matters in practice because shorter component series provide less information and therefore should generally receive less effective influence in the forecast aggregation.
25.4 Seasonal Decomposition Without Parametric Filters
A major strength of the NNS approach is that seasonality is handled directly through component decomposition rather than through fixed harmonic terms or pre-imposed smoothing filters.
Classical methods often represent seasonality through:
- seasonal ARIMA operators,
- trigonometric terms,
- moving-average filters,
- or exponential smoothing recursions.
The NNS approach instead asks a simpler question:
At which lag lengths does the series become more predictable when split into component sequences?
This is operationalized through a seasonality test based on the coefficient of variation of each component series relative to the coefficient of variation of the full series.
If a lag \(m\) produces component series with lower coefficient of variation than the original series, then the lag reveals recurring structure. Intuitively:
- a lower component-series coefficient of variation means tighter local behavior,
- tighter local behavior means greater predictability,
- and greater predictability indicates seasonality or cyclic structure.
Thus seasonality is not defined through a parametric frequency-domain object. It is defined through predictive concentration in lag-defined subsets.
25.5 Seasonal Detection by Predictive Power
Let the full series have coefficient of variation
\[ CV(X) = \frac{\sigma_X}{|\mu_X|}, \]
assuming the mean is nonzero.
For a candidate lag \(m\), construct the \(m\) component series. Let their representative predictive concentration be summarized through their component coefficients of variation.
If these component coefficients are systematically lower than the overall series coefficient of variation, then the lag \(m\) is informative.
The interpretation is immediate:
- lower \(CV\) means less dispersion relative to level,
- less dispersion means more stable phase-specific structure,
- more stable phase-specific structure means improved forecastability.
This gives a nonparametric test for seasonality grounded in prediction rather than in harmonic decomposition.
The chapter’s conceptual point is broader than the specific diagnostic:
seasonality is treated as recurring conditional structure, not as a parametric periodic law.
25.5.1 A worked illustration
Consider the toy quarterly series
\[ X = (10, 18, 11, 21, 12, 20, 13, 23). \]
The overall mean is
\[ \bar X = \frac{10+18+11+21+12+20+13+23}{8} = 16, \]
and the sample standard deviation is approximately
\[ s_X \approx 5.071. \]
Hence the overall coefficient of variation is
\[ CV(X) \approx \frac{5.071}{16} = 0.317. \]
Now test lag \(m=2\). The component series are
\[ (10,11,12,13) \quad\text{and}\quad (18,21,20,23). \]
For the first component,
\[ \bar X_1 = 11.5,\qquad s_1 \approx 1.291,\qquad CV_1 \approx \frac{1.291}{11.5}=0.112. \]
For the second,
\[ \bar X_2 = 20.5,\qquad s_2 \approx 2.082,\qquad CV_2 \approx \frac{2.082}{20.5}=0.102. \]
Both component coefficients of variation are far below the overall value \(0.317\). That means the lag-2 decomposition produces tighter, more internally coherent subseries than the original series. In the NNS interpretation, lag \(2\) reveals meaningful recurring structure.
By contrast, consider lag \(m=3\). The three component series are
\[ (10,21,13),\qquad (18,12,23),\qquad (11,20). \]
These exhibit much larger within-component variation, so their component coefficients of variation are not uniformly smaller than the full-series value. In this case lag \(3\) is not as predictive as lag \(2\).
This simple example shows exactly how the test works in practice. One computes the overall coefficient of variation, computes the component-series coefficients of variation for each candidate lag, and prefers those lags for which the component series become materially tighter than the original sequence.
Mathematically, the logic is straightforward: if
\[ CV_j(m) < CV(X) \quad \text{for component series } j=1,\dots,m, \]
then conditioning on phase within lag \(m\) reduces relative dispersion. Reduced relative dispersion means more concentrated conditional behavior, and more concentrated conditional behavior implies improved forecastability.
In applications one usually needs an operational aggregation rule. A natural choice is a weighted average of component coefficients of variation,
\[ \overline{CV}(m) = \sum_{j=1}^{m} w_j\,CV_j(m), \qquad w_j \ge 0, \qquad \sum_{j=1}^{m} w_j = 1, \]
where \(w_j\) may be proportional to component length or to some predictive-strength measure. Then lag \(m\) is favored when
\[ \overline{CV}(m) < CV(X). \]
An even stricter rule requires most, or all, component CVs to lie below the full-series CV. The exact rule is a modeling choice, but the guiding principle is unchanged: a good seasonal lag is one that produces tighter conditional distributions than the unpartitioned series.
Component length also matters. The sample standard deviation of \(k\) observations is biased low, by a factor that depends on the distribution (\(c_4(2) \approx 0.80\) for normal data, less for skewed or heavy-tailed data), so at very long lags, where each component has only two or three points, every component looks tight even when the series is pure noise. NNS.seas therefore measures dispersion with the Gini mean difference, the degree-one lower plus upper partial moment about each observation in turn,
\[ G(X) = \frac{1}{k(k-1)} \sum_{i \ne j} |X_i - X_j|, \]
which is unbiased for the population mean pairwise absolute difference under iid sampling with a finite first moment. Serial dependence changes that expectation, so this is not an unconditional unbiasedness claim for every time series. The scan retains only lags with at least five observations in every component (\(m \le n/5\) by default, or fewer through max.period). The eight-point series above is a worked illustration below that size; NNS.seas needs at least ten observations to test lag 2.
A trend can inflate component dispersion and hide a period, so NNS.seas removes the linear trend first while keeping the mean level. Noise can still make components appear more concentrated by chance. Detected periods are therefore candidates for forecasting validation. With p.value = TRUE, each ratio is compared with the best ratio over all lags in shuffled copies. This accounts for the scan under an exchangeable residual null; it is not a generally calibrated no-seasonality test for arbitrary serial dependence. For dependent nonseasonal data, calibrate an appropriate null separately and assess forecast performance at new chronological origins.
A technical caveat is also important. The coefficient of variation needs a ratio scale: if the series crosses zero, \(\mu_X\) can be near zero and \(CV(X)\) unstable, and the number of lags that pass would depend on the level of the series rather than on its structure. NNS.seas therefore uses the coefficient of variation only when the series stays on one side of zero, and compares the component dispersions directly otherwise. The predictive logic is the same either way. For an interval-scale series, where adding a constant is only a change of units or origin, use dispersion = "absolute". It centers the series before the same Gini-dispersion scan and makes the score invariant to additive offsets. The default dispersion = "auto" retains the original ratio-scale rule.
25.6 Multiple Seasonalities
Many real series have more than one recurring period.
Examples include:
- monthly data with annual and multi-year cycles,
- hourly data with daily and weekly cycles,
- sales data with weekly, monthly, and promotional rhythms.
Classical models often struggle when multiple seasonalities interact, especially when the interactions are nonlinear or when the seasonal periods are not cleanly nested.
The NNS framework handles this naturally by allowing multiple candidate lags:
\[ m_1, m_2, \dots, m_K. \]
Each lag defines its own component decomposition and its own forecast. These lag-specific forecasts can then be combined using weights reflecting both:
- the predictive tightness of the corresponding component series,
- and the amount of information available within each lag structure.
This is a central advantage of the nonparametric formulation: multiple seasonal patterns need not be forced into one rigid dynamic equation. They can be estimated as separate predictive structures and then aggregated.
25.7 Predictive Lag and Fundamental Period
A multiple of the true period scores as well in expectation: the components of lag \(2m\) are every other observation of the components of lag \(m\). Its components are shorter, though, so its score is noisier, and the minimum over many near-equal scores tends to fall on a multiple. For a monthly sine with noise (\(n = 540\)), the lowest-scoring lag was a multiple of 12, between 36 and 108, in all of 40 simulated series. NNS.seas therefore reports the fundamental period as best.period: the smallest period that divides the lowest-scoring lag (or nearly divides it, allowing for a non-integer period such as 365.25 days) and whose score exceeds the lowest by at most two jackknife standard errors of their difference, computed from paired delete-one scores because both scores reuse every observation. On those series it returns 12 in every case, as it does with \(t_3\) noise and in 37 of 40 series with exponential noise, while a true 24-month pattern stays at 24. nottem moves from 24 to 12 and lynx from 19 to its 10-year cycle. Over rolling forecast origins on these and other series, with the period re-selected from the training data at each origin, the NNS.ARMA forecasts at the fundamental period had the lower mean MASE wherever the period changed (1.05 against 1.20 over 52 fits).
The fundamental period is listed first in all.periods and periods, and the lowest-scoring lag remains available from the scores in all.periods. The minimum of five observations per phase limits the largest tested period; max.period and modulo expose that search domain and should be reported with the result. Validate selected predictive lags only on data available before each forecast origin.
NNS.ARMA(seasonal.factor = TRUE) uses best.period without the permutation test, which would add nperm scans to every call. For a significance-filtered lag, run the two steps explicitly: NNS.seas(x, p.value = TRUE) first, then pass its best.period (1 when no period is significant) or periods as seasonal.factor.
25.8 Nonlinear Autoregressive Structures
The classical AR model is linear in lagged values. NNS replaces this with a more general dynamic relation.
If \(X_t\) depends on past observations through a nonlinear rule, then there is no reason to insist that the forecast be generated by a straight line fitted to lagged values.
For a phase \(j\) and candidate period \(m\), the implemented component regression has the form
\[ X_{j+km}=g_{j,m}(k)+\varepsilon_{j,k}. \]
It regresses component values against their occurrence index \(k\), then extrapolates that phase. This differs from directly fitting \(X_t=f(X_{t-m})\) to pairs of lagged values. Both are informed by the series’ own history, but only the former describes the current NNS.ARMA component engine. The nonlinear \(g_{j,m}\) uses the regression machinery from earlier chapters and can capture:
- turning points,
- diminishing effects,
- local curvature,
- asymmetric phase behavior,
- and regime-like transitions.
The importance of this step cannot be overstated.
A component series may itself be nonlinear even when the original series looks smooth. If the local phase-specific dynamics are nonlinear, a linear subseries regression can point in the wrong direction entirely. Nonparametric regression is therefore not a cosmetic addition. It is the mechanism that allows autoregression to remain autoregressive without remaining linear.
25.9 Directional Temporal Dependence
Time-series dependence is not merely contemporaneous dependence shifted through time. It has direction:
- earlier values may help predict later values,
- later values cannot influence earlier ones,
- and positive versus negative deviations may propagate differently across time.
Within the broader NNS framework, this suggests a temporal analogue of the co-partial-moment decomposition developed in Chapters 9–10 and 14. For a lag \(\tau \ge 1\), one can study lagged co-partial moments formed from aligned pairs such as \((X_{t-\tau}, X_t)\), separating concordant movement from divergent movement across time.
In that interpretation, one class of lagged moments captures persistence in the same directional regime, while another captures reversals between periods. The distinction is useful because many time series are dynamically asymmetric even when their unconditional summaries appear mild.
For example:
- volatility clusters after large shocks,
- downturns may persist longer than upswings,
- inventory shortages may propagate differently than surpluses,
- and demand spikes may reverse more sharply than demand collapses.
Directional temporal dependence therefore generalizes classical autocorrelation by preserving regime-specific information that linear autocovariance averages away.
It is important, however, to place this idea correctly within the NNS framework. In the current univariate forecasting routines, time dependence is operationalized primarily through lag-defined component regression and seasonality detection, not through a standalone directional-autocorrelation statistic. The lagged co-partial-moment construction belongs most naturally to the theory developed for asymmetric dependence and causation, where temporal ordering is analyzed explicitly rather than only through forecast generation. Readers can map this to Chapter 10 for directional dependence, Chapter 12 for the copula interpretation, and Chapter 14 for directional-causation asymmetry.
25.10 Forecasting from Component Regressions
The NNS forecasting workflow can now be stated clearly.
25.10.1 Step 1: Select candidate seasonal lags
Identify one or more plausible lag lengths, either from domain knowledge or from the predictive seasonality test.
25.10.2 Step 2: Form component series
For each lag \(m\), partition the original series into \(m\) phase-specific subseries.
25.10.3 Step 3: Regress each component forward
For each component series, estimate the next observation using either:
- linear regression,
- nonlinear NNS regression,
- both, or
- mean-based shrinkage variants.
25.10.4 Step 4: Weight and aggregate
Combine component forecasts using weights that reflect predictive concentration and sample support.
25.10.5 Step 5: Iterate if forecasting multiple steps ahead
For multi-step forecasting, append the newly predicted value and repeat. Seasonal factors may be kept fixed or updated dynamically as the forecast path evolves.
This procedure preserves the definition of autoregression:
the forecast is still generated from the series’ own past.
But it does so without imposing stationarity, without requiring Box–Jenkins identification, and without restricting the lag relation to a linear map.
A brief clarification of the mean-based option is useful. In some component series the fitted regression may be unstable because the component is short, noisy, or nearly flat. In that case a practical alternative is to shrink the regression estimate toward the component mean, or even to use the component mean directly. This sacrifices some responsiveness in exchange for stability. Conceptually, it is a local bias-variance tradeoff: when the estimated slope or nonlinear fit is unreliable, the component average can act as a robust anchor.
# Univariate nonlinear ARMA: a sine of period 2 * pi * 8 (about 50) in noise
set.seed(123)
raw_series <- sin(seq_len(480) / 8) + rnorm(480, sd = .35)
# Estimate scaling and select periods using only the training prefix.
validation_origin <- floor(.8 * length(raw_series))
training <- raw_series[seq_len(validation_origin)]
z <- (raw_series - mean(training)) / sd(training)
# z crosses zero, so component dispersions are compared directly.
seasonal_period <- NNS.seas(z[seq_len(validation_origin)], plot = FALSE)
head(seasonal_period$all.periods)
# Validate seasonal periods and forecast
NNS.ARMA.optim(z, h = 48, training.set = validation_origin,
seasonal.factor = seasonal_period$periods, plot = TRUE, ncores = 1)The same workflow in Python:
import numpy as np
from nns import nns_seas, nns_arma_optim
# Univariate nonlinear ARMA
rng = np.random.default_rng(123)
raw_series = np.sin(np.arange(1, 481) / 8) + rng.normal(scale=0.35, size=480)
# Estimate scaling and select periods from the training prefix
validation_origin = int(.8 * len(raw_series))
training = raw_series[:validation_origin]
z = (raw_series - training.mean()) / training.std(ddof=1)
seasonal_period = nns_seas(z[:validation_origin], plot=False)
seasonal_period["all.periods"]
# Validate seasonal periods and forecast
nns_arma_optim(z, h=48, training_set=validation_origin,
seasonal_factor=seasonal_period["periods"])The figure below was regenerated with the current seasonality, recursion, and
validation rules. The complete executable R source is
book/scripts/ch25_sine_example.R. From the repository root, run
Rscript book/scripts/ch25_sine_example.R book/images/ch25_uni_ts.png to
regenerate the result and figure; omitting the output path prints the result only.
The script also records sessionInfo(). The period scan and all preprocessing
use the training prefix; a whole-series standardization would use later values.
For this zero-crossing series, the historical coefficient-of-variation column
names contain Gini dispersions. Standardizing the prefix’s standard deviation
to one does not make its Gini dispersion one. The optimizer’s score is internal
validation, not an independent final test.
25.11 Dynamic Updating and Recursive Forecast Paths
A one-step forecast is rarely the end goal. In practice, analysts often require
\[ h = 1,2,\dots,H \]
steps ahead.
In the NNS framework, multi-step forecasting proceeds recursively.
If \(\hat X_{T+1}\) is forecast first, then it is appended to the series and treated as part of the evolving path when forecasting \(\hat X_{T+2}\), and so on.
This creates two natural modes.
25.11.1 Static seasonal structure
The seasonal lags and weights are estimated once from the historical sample and then held fixed for all future steps.
25.11.2 Dynamic seasonal structure
The seasonal structure is recomputed as the forecast path grows, allowing the decomposition itself to evolve.
The static approach favors stability. The dynamic approach favors adaptability.
But that distinction can be made more precise.
Static updating is generally preferable when:
- the dominant seasonal pattern is well known in advance,
- the series is long enough that seasonal weights are already stable,
- the forecast horizon is short relative to the seasonal period,
- or the analyst values interpretability and reproducibility over rapid adaptation.
In such cases, recomputing the lag structure at every step may add noise rather than information. If the underlying periodic structure is persistent, then a fixed decomposition acts as a stabilizer.
Dynamic recomputation is preferable when the data suggest that the seasonal structure itself is moving. Typical signals include:
- abrupt level shifts,
- changing amplitudes of recurring cycles,
- newly emerging or fading periodicities,
- strong structural breaks,
- or forecast errors that begin to cluster by phase.
For example, a retail series may historically be dominated by an annual pattern, yet after a major platform change or supply shock, shorter promotional cycles may become more predictive than the old annual rhythm. In that case, holding the original seasonal factor fixed can lock the forecast into an outdated regime. Dynamic updating allows the decomposition to respond as the series evolves.
So the practical decision is not merely “stability versus adaptability.” It is a question of whether the analyst believes the lag structure is itself part of the stable signal or part of the changing environment.
A useful empirical guide is out-of-sample validation. One may reserve a holdout period, compare static and dynamic forecasts over that window, and select the updating rule that yields better predictive accuracy. In that sense, the static-versus-dynamic decision is not purely philosophical. It can be treated as a forecasting design choice subject to cross-validation.
25.12 Bias Correction and Its Validation
NNS.ARMA.optim retains the gravity-based residual shift as an option, but does not apply it by default (bias.correction = "none"). Accepting a shift because it improves the objective on the window it was estimated from is not evidence that the bias persists, and on rolling-origin forecasts of nine series (six expanding origins, h = 12) the shifted forecasts had the higher mean MASE on eight. bias.correction = "split" estimates the shift from the first half of the internal validation forecasts, then accepts it only when the objective improves on the later half; that separation removed the same-window acceptance but did not make the shift carry over to the forecast window on those series. The support rule available at the validation origin is applied before scoring the candidate correction. A later final refit may update that rule from the full history then available; validation must not use later negative observations to relax a nonnegative training support. At least four validation observations are needed; shorter windows leave the shift disabled.
The returned bias.validation records the estimation and acceptance rows, candidate and applied shifts, scores and decision. raw.results retains the forecast before the shift, and validation.forecast identifies the forecast used for the final internal objective. Earlier versions estimated and accepted the shift on the same validation window; that rule nearly always passes its own test and is no longer offered.
This separation tests whether an estimated bias persists later in time. It does not make the whole fit independently validated: seasonal factors, methods, shrinkage and smoothing still use the internal validation window. Compare raw and corrected complete forecasts at fresh rolling origins, and evaluate coverage separately from the point-forecast objective.
25.13 Prediction Intervals for Forecasts
Point forecasts are only one part of the forecasting problem. Analysts also need measures of uncertainty.
The direct NNS.ARMA forecast bands use the maximum entropy bootstrap machinery from Chapter 20. Directional quantiles avoid fitting a Gaussian error law, although the bootstrap defaults include the iid normal-reference adjustment of replicate means described there.
NNS.ARMA(..., pred.int = ) applies it to the forecast path itself. Suppose the fitted model yields
\[ \hat X_{T+1}, \dots, \hat X_{T+H}. \]
The procedure is:
- generate scenarios from this path with the maximum entropy bootstrap across dependence targets \(\rho \in [-1,1]\), varying its residual rank structure;
- at each step, take the directional quantiles \(\text{LPM.VaR}(\alpha/2)\) and \(\text{UPM.VaR}(\alpha/2)\) of the replicates;
- widen them by the spread of the per-lag linear estimates.
Because the replicates are generated from the \(H\) forecast values, the resulting band describes variation of synthetic forecast paths, not a separately estimated future-error distribution at each horizon. The construction has no calibrated rule for accumulating forecast error with the horizon: seasonal paths can give wide bands and flat paths narrow ones. A nominal label of \(1-\alpha\) therefore does not establish that fraction of future outcomes will be covered. Measure coverage at new chronological origins, by horizon and data-generating setting.
A predictive interval in the sense of Chapter 17, \(P(\text{LPM.VaR}(\alpha/2) \le X_{T+h} \le \text{UPM.VaR}(\alpha/2)) \approx 1 - \alpha\), needs the distribution of future outcomes at each horizon instead. Two routes stay within the framework: directional quantiles of validation errors collected separately for each horizon, which NNS.ARMA.optim approximates with its validation-window errors, or synthetic continuations of the observed series, rerunning the recursion on each bootstrap replicate and taking directional quantiles of the simulated outcomes at each horizon. Whichever is used, the interval calibration check in the validation protocol below applies.
25.14 Relation to Earlier NNS Chapters
Time-series modeling in NNS is not an isolated technique. It is a direct extension of earlier ideas.
25.14.2 From Chapter 19
The local regression on component series inherits the data-adaptive bandwidth logic of partition estimators.
25.15 Comparison with ARIMA
ARIMA remains one of the benchmark tools of time-series forecasting.
Its strengths are well known:
- interpretable lag operators,
- strong theory under stationarity,
- effective performance on linear stochastic dynamics.
But its limitations are equally clear when viewed through the directional lens.
25.15.1 Structural form
ARIMA assumes a linear dependence structure after differencing. NNS does not.
25.15.2 Stationarity
ARIMA is built around stationarity and invertibility conditions. NNS forecasting does not require the series to satisfy a stationary parametric model in level.
25.15.3 Identification burden
ARIMA requires order selection and specification diagnostics. NNS shifts the task from parametric identification to predictive lag decomposition.
25.15.4 Nonlinearity
ARIMA can approximate some nonlinear behavior through transformations or hybridization, but nonlinearity is not native to the model. In NNS, it is native.
This does not mean ARIMA is obsolete. It means that ARIMA is best understood as a special, linear, tightly structured case of a broader forecasting problem.
For balance, however, NNS does not eliminate modeling choices. It replaces ARIMA’s order-identification problem with choices about lag selection, regression method, aggregation weights, and updating scheme. The difference is methodological rather than absolute: in NNS, these choices are naturally evaluated by predictive performance rather than by adherence to a pre-specified parametric identification protocol.
25.16 Comparison with ETS Models
ETS methods model time series through combinations of
- error,
- trend,
- and seasonality,
typically using exponential smoothing recursions and state-space interpretations.
These methods are often highly effective, especially on business forecasting problems with stable level, trend, and seasonal components.
Relative to ETS, the NNS approach differs in several ways.
25.16.1 Component meaning
ETS decomposes the series into latent level, trend, and seasonality states. NNS decomposes it into lag-defined predictive subsets.
25.16.2 Smoothing mechanism
ETS uses recursive smoothing equations. NNS uses local regression and weighted aggregation across component series.
25.16.3 Parametric structure
ETS specifies an updating architecture in advance. NNS lets the predictive structure emerge from the data.
25.16.4 Nonlinear interactions
ETS can adapt smoothly, but it is not inherently designed for rich nonlinear autoregressive geometry. NNS is.
The practical difference is conceptual:
ETS smooths a presumed component architecture. NNS learns a predictive architecture from subset behavior.
A further distinction is distributional. Many ETS formulations are estimated in likelihood-based frameworks tied to specific error models, often Gaussian or close variants. The NNS approach imposes no such distributional law on the series or on the forecast errors.
25.17 When the NNS Approach Is Especially Useful
The nonparametric time-series framework is particularly attractive when one or more of the following hold:
- the series exhibits nonlinear cyclic behavior,
- multiple seasonalities are present,
- model stationarity is doubtful,
- lag effects are structurally asymmetric,
- parametric identification is fragile,
- or prediction accuracy matters more than adherence to a classical stochastic specification.
Examples include:
- retail and transaction flows,
- cyclical economic indicators,
- energy demand,
- financial and commodity time series,
- and operational processes with threshold-driven dynamics.
These are precisely the settings where local structure matters more than global parametric elegance.
25.18 Limitations
A nonparametric forecasting method is not free of tradeoffs.
25.18.1 Primarily univariate in this chapter
The methods developed here focus on a single series. Cross-series interactions are deferred to Chapter 26.
25.18.2 Data requirements
Because prediction is learned from historical structure, sparse component series limit reliability for very large lag lengths. The default NNS.seas scan requires at least five observations per component. Explicit modulo candidates can extend that search; they need separate predictive validation and do not acquire the support of the default scan simply by being supplied.
25.18.3 Computational cost
Searching many seasonal combinations and fitting nonlinear regressions can be more computationally intensive than fitting a simple linear ARIMA.
25.18.4 Interpretability of dynamics
A fitted ARIMA equation provides direct coefficients. NNS instead provides a predictive mechanism based on component regressions and weights. This is often more flexible, but less compact as a closed-form law.
These limitations are real. But they are the cost of avoiding the stronger assumptions of parametric time-series models.
The multivariate extension is natural in principle, but not automatic in implementation. Once several series enter, one must distinguish self-dependence from cross-dependence, align potentially different frequencies, and account for lead-lag structure across variables.
25.19 Leakage-Safe Backtesting Protocol
Forecasting performance must be assessed with strict time-order preservation. A leakage-safe protocol is:
- Rolling-origin evaluation: choose an initial training window
[1, T_0]. - Forecast horizon definition: for each origin
t, produce forecasts fort+hwithout using observations aftert. - Expanding or sliding refit:
- expanding window: train on
[1, t], or - sliding window: train on
[t-w+1, t].
- expanding window: train on
- No future-informed preprocessing: any scaling, interpolation, imputation, or feature construction must be computed using data available at origin
tonly. - Horizon-specific scoring: report MAE/RMSE/coverage separately for each horizon
hrather than pooling all horizons. - Interval calibration check: compare nominal vs empirical coverage for prediction intervals across horizons.
For seasonal component construction, lag selection must also be origin-specific; selecting global lags from the full sample before backtesting constitutes leakage.
25.20 Summary
This chapter developed nonparametric time-series modeling in the NNS framework.
Its main contributions are fivefold.
First, it reframed time series as a subset regression problem. Forecasting is treated as conditional estimation on lag-defined component series rather than as fitting a single global linear recursion.
Second, it developed seasonal decomposition by predictive concentration. Seasonal structure is detected through reductions in coefficient of variation across component series, linking seasonality directly to forecastability.
Third, it established nonlinear autoregressive forecasting. Component series are projected forward using nonparametric regression, allowing local curvature and asymmetric dynamics to enter the forecast natively.
Fourth, it clarified how the book’s broader dependence framework extends into time. Lagged directional structure can be studied theoretically through co-partial-moment ideas, even though the chapter’s main forecasting routines operationalize time dependence through component regression and seasonality rather than through a standalone directional-autocorrelation statistic.
Fifth, it clarified the chapter’s relationship to classical methods. ARIMA and ETS remain important special-purpose tools, but they impose structural assumptions that the NNS framework avoids.
Taken together, these results show that the directional nonparametric framework extends naturally from cross-sectional estimation to temporal prediction.
But the framework developed here remains fundamentally univariate. Once other series matter, the main difficulty is no longer just whether the past of \(X_t\) predicts its future, but whether lagged values of \(Y_t\), \(Z_t\), and other related processes alter that forecast in nonlinear and asymmetric ways. In that setting, univariate decomposition can miss cross-series lead-lag effects, common shocks, and mixed-frequency structure.
The next chapter therefore generalizes the same ideas to multivariate forecasting, where multiple time series interact through directional dependence, lagged cross-variable structure, and mixed-frequency information.
Further Reading / Examples
For forecasting applications, including the tidal data example, see the NNS Time-Series Forecasting Examples. This behavior is illustrated in the historical tidal forecasting example. Its archived output has not been regenerated under the current validation protocol; see the reproducibility notes before drawing numerical conclusions.