CLAIMSET Complex exponential smoothing summary Exponential smoothing forecasts a series by first choosing a model: with or without a trend, with or without seasonality. This paper proposes complex exponential smoothing (CES), which uses complex variables, pairing each value with its forecast error, so that one model moves smoothly between a level and a trend as its smoothing parameter changes. It shows that CES can be stationary or non-stationary depending on that parameter, and that it corresponds to an ARMA(2,2) model whose coefficients are tied together. A seasonal version handles seasonality without seasonal indices, so model selection reduces to a choice between two models. On the series of the M1, M3, tourism and M4 forecasting competitions, CES was more accurate than exponential smoothing, ARIMA and the Theta method on most error measures, though not on all, and adding it to a combination of forecasts improved the combination. licence CC-BY-4.0; contributions C1, C2, C3, C4, C5, C6, C7, C8, C9, C10, C11, C12; highlights D1, C2, C10, C4, C12 declarations funding: none | competing interests: None declared. DEFINITIONS D1 [method] CES: Complex exponential smoothing: a forecasting method that combines each past observation $y_t$ and its forecast error $e_t$ into the complex variable $y_t + i e_t$ and forecasts with an exponentially weighted sum of these values, $$\hat{y}_t + i\hat{e}_t = (\alpha_0 + i\alpha_1) \sum_{j=1}^{t-1} (1 + i - (\alpha_0 + i\alpha_1))^{j-1} (y_{t-j} + i e_{t-j})$$, or recursively $$\hat{y}_t + i\hat{e}_t = (\alpha_0 + i\alpha_1)(y_{t-1} + i e_{t-1}) + (1 - \alpha_0 + i - i\alpha_1)(\hat{y}_{t-1} + i\hat{e}_{t-1})$$, where $\alpha_0 + i\alpha_1$ is the [[D2:complex smoothing parameter]] and $\hat{e}_t$ carries the information on past forecast errors. [assumes C60, C62, C63, C64] D2 [variable] complex smoothing parameter: The smoothing parameter of [[D1:CES]], the complex number $\alpha_0 + i\alpha_1$; its real part $\alpha_0$ and imaginary part $\alpha_1$ are the two real parameters estimated, and they set how much of the past observations and of the past forecast errors is carried into the forecast. [part of D1] D3 [concept] nonlinear trend component: The state $c_t$ of the [[D1:CES]] state-space model, which carries information on both its own previous value $c_{t-1}$ and the previous level $l_{t-1}$; it does not correspond to the trend component of exponential smoothing. [part of D1] D4 [method] seasonal CES: [[D1:CES]] extended with a second pair of level and [[D3:nonlinear trend]] states taken at lag $m$, the seasonal period, with its own [[D2:complex smoothing parameter]] $\beta_0 + i\beta_1$: $y_t = l_{0,t-1} + l_{1,t-m} + \varepsilon_t$, $$l_{0,t} = l_{0,t-1} - (1 - \alpha_1) c_{0,t-1} + (\alpha_0 - \alpha_1) \varepsilon_t$$, $$c_{0,t} = l_{0,t-1} + (1 - \alpha_0) c_{0,t-1} + (\alpha_0 + \alpha_1) \varepsilon_t$$, $$l_{1,t} = l_{1,t-m} - (1 - \beta_1) c_{1,t-m} + (\beta_0 - \beta_1) \varepsilon_t$$, $$c_{1,t} = l_{1,t-m} + (1 - \beta_0) c_{1,t-m} + (\beta_0 + \beta_1) \varepsilon_t$$. [a kind of D1] D5 [method] SCUM: Simple combination of univariate models: the median of the forecasts of [[D9:ETS]], [[D10:ARIMA]], the [[D11:Theta method]] and [[D17:CES]]. [source B13] D6 [method] SCUM(noCES): The median of the forecasts of [[D9:ETS]], [[D10:ARIMA]] and the [[D11:Theta method]]: [[D5:SCUM]] without CES. [a kind of D5] D7 [method] the six compared methods: The [[D12:four compared methods]] and the two combinations, [[D5:SCUM]] and [[D6:SCUM(noCES)]]. [members D17, D9, D10, D11, D5, D6] D8 [data] M1, M3 and tourism data: The 5,315 yearly, quarterly, monthly and other time series of the M1, M3 and tourism forecasting competitions. D9 [method] ETS: Exponential smoothing in a state space framework with automatic selection among its models; here es() from the smooth package with model ZXZ, selecting by AICc. [source B2] D10 [method] ARIMA: Automatic state space ARIMA; here auto.ssarima() from the smooth package. [source B15] D11 [method] Theta method: The Theta method of forecasting; here thetaf() from the forecast package 8.14. [source B16] D12 [method] the four compared methods: [[D17:CES with automatic seasonal selection]], [[D9:ETS]], [[D10:ARIMA]] and the [[D11:Theta method]], as the benchmark runs them. [members D17, D9, D10, D11] D13 [method] the three benchmark methods: [[D9:ETS]], [[D10:ARIMA]] and the [[D11:Theta method]], as the benchmark runs them. [members D9, D10, D11] D14 [variable] RMSSE: Root mean squared scaled error: the root mean squared forecast error over the horizon, scaled by the mean absolute first difference of the in-sample data; reported as its mean and median over the series. [source B17] D15 [variable] MASE: Mean absolute scaled error: the mean absolute forecast error over the horizon, scaled by the mean absolute first difference of the in-sample data; reported as its mean and median over the series. [source B18] D16 [variable] MSIS: Mean scaled interval score: the interval score of a prediction interval averaged over the horizon and scaled by the mean absolute first difference of the in-sample data; here of the 95% prediction interval, reported as its mean and median over the series. [source B19] D17 [method] CES with automatic seasonal selection: [[D1:CES]] fitted by auto.ces() from the smooth package 3.1.2, which chooses between [[D4:seasonal CES]] and non-seasonal CES by AICc. [a kind of D1] D18 [method] non-seasonal special cases of ETS: The ETS models [[D19:ETS(A,N,N)]], [[D20:ETS(A,A,N)]], [[D21:ETS(A,Ad,N)]], [[D22:ETS(M,M,N)]] and [[D23:ETS(M,Md,N)]], each fitted without model selection. [members D19, D20, D21, D22, D23] [source B1] D19 [method] ETS(A,N,N): The ETS model with additive error, no trend and no seasonality, fitted without model selection. [source B1] D20 [method] ETS(A,A,N): The ETS model with additive error, additive trend and no seasonality, fitted without model selection. [source B1] D21 [method] ETS(A,Ad,N): The ETS model with additive error, additive damped trend and no seasonality, fitted without model selection. [source B1] D22 [method] ETS(M,M,N): The ETS model with multiplicative error, multiplicative trend and no seasonality, fitted without model selection. [source B1] D23 [method] ETS(M,Md,N): The ETS model with multiplicative error, multiplicative damped trend and no seasonality, fitted without model selection. [source B1] D24 [method] seasonal special cases of ETS: The ETS models [[D27:ETS(A,N,A)]], [[D28:ETS(A,A,A)]], [[D29:ETS(A,Ad,A)]], [[D30:ETS(M,M,M)]] and [[D31:ETS(M,Md,M)]], each fitted without model selection. [members D27, D28, D29, D30, D31] [source B1] D25 [method] CES and the non-seasonal special cases of ETS: [[D17:CES with automatic seasonal selection]] and the [[D18:non-seasonal special cases of ETS]], as compared on the [[D33:non-seasonal subset]]. [members D17, D19, D20, D21, D22, D23] D26 [method] CES and the seasonal special cases of ETS: [[D17:CES with automatic seasonal selection]] and the [[D24:seasonal special cases of ETS]], as compared on the [[D34:seasonal subset]]. [members D17, D27, D28, D29, D30, D31] D27 [method] ETS(A,N,A): The ETS model with additive error, no trend and additive seasonality, fitted without model selection. [source B1] D28 [method] ETS(A,A,A): The ETS model with additive error, additive trend and additive seasonality, fitted without model selection. [source B1] D29 [method] ETS(A,Ad,A): The ETS model with additive error, additive damped trend and additive seasonality, fitted without model selection. [source B1] D30 [method] ETS(M,M,M): The ETS model with multiplicative error, multiplicative trend and multiplicative seasonality, fitted without model selection. [source B1] D31 [method] ETS(M,Md,M): The ETS model with multiplicative error, multiplicative damped trend and multiplicative seasonality, fitted without model selection. [source B1] D32 [data] M4 data without 10 series: The 100,000 series of the M4 competition (23,000 yearly, 24,000 quarterly, 48,000 monthly, 359 weekly, 4,227 daily and 414 hourly) except series 19357, 43160, 51552, 62738, 92370, 92435, 93159, 93196, 93412 and 93491, whose influential outliers and structural breaks made ETS break down. [source B9] D33 [data] non-seasonal subset: The 3,262 series of the [[D8:M1, M3 and tourism data]] for which ETS model selection by es() from the smooth package chooses a non-seasonal model. [part of D8] [assumes C65] D34 [data] seasonal subset: The 2,053 series of the [[D8:M1, M3 and tourism data]] for which ETS model selection by es() from the smooth package chooses a seasonal model. [part of D8] [assumes C65] D35 [data] the seasonal and non-seasonal subsets: The [[D33:non-seasonal subset]] and the [[D34:seasonal subset]] of the [[D8:M1, M3 and tourism data]]. [members D33, D34] D36 [design] holdout at the M1, M3 and tourism horizons: Each series split into an in-sample part and a holdout of the length used in its competition: 6 observations for yearly, 8 for quarterly and other, and 18 for monthly series; forecasts made from the end of the in-sample part. D37 [design] holdout at the M4 horizons: Each series split into an in-sample part and a holdout of the length used in the M4 competition: 6 observations for yearly, 8 for quarterly, 18 for monthly, 13 for weekly, 14 for daily and 48 for hourly series; forecasts made from the end of the in-sample part. D38 [design] Nemenyi test: Post hoc Nemenyi test on the ranks of an error measure across series, comparing all methods at the 5% level; here rmcb() from the greybox package. [source B20] D39 [method] Akaike information criterion: A criterion for choosing among models fitted by maximum likelihood that penalises the maximised log-likelihood by the number of estimated parameters. [source B21] D40 [design] ADF test: Augmented Dickey-Fuller test of the null hypothesis that a time series has a unit root. [source B22] D41 [design] KPSS test: Kwiatkowski-Phillips-Schmidt-Shin test of the null hypothesis that a time series is stationary. [source B23] D42 [data] M3 series 1664 and 2721: Two monthly series of the M3 competition: series 1664, which graphical analysis and the [[D40:ADF]] and [[D41:KPSS tests]] indicate is stationary, and the trended series 2721; each split into in-sample and holdout parts. CLAIMS C1 [formal; principal; extension; support: formal_proof E3; uses C17, C18, C19, C20, C23, C24, C25, C32, C21, C22, C33; verification unchecked; verifiability open] Depending on the value of its [[D2:complex smoothing parameter]], [[D1:CES]] is a stationary or a non-stationary model. extends B1, own earlier work: Exponential smoothing in a state space framework (ETS) gives 30 models with no, additive or multiplicative error, trend and seasonal components, among which a model is selected by an information criterion. addresses B3, own earlier work: Identifying the presence and type of trend in exponential smoothing often leads to misidentification and poorly performing long-term forecasts. addresses B2: The diverse trend alternatives of exponential smoothing in a state space framework make the selection of a model difficult. addresses B4: Selecting the most appropriate exponential smoothing model is a challenging task. addresses B5, own earlier work: Even when the components of exponential smoothing are identified correctly, the estimated smoothing parameters and the weights they imply can give poor forecasts. C2 [formal; principal; support: formal_proof E3; uses C17, C18, C25; verification unchecked; verifiability open] [[D1:CES]] passes continuously from a stationary to a non-stationary model as its [[D2:complex smoothing parameter]] changes. C3 [descriptive; principal; support: empirical_quant E2 (E2.T1.r1, E2.T1.r2, E2.T1.r3, E2.T1.r4); uses C34; limitations L1, L3; inherits C60, C62, C63, C64; verifiability open] On the [[D8:M1, M3 and tourism data]], [[D17:CES]] has the lowest mean [[D14:RMSSE]] and the lowest median RMSSE, [[D15:MASE]] and [[D16:MSIS]] of [[D12:four methods, CES, ETS, ARIMA and the Theta method]], but not the lowest mean MASE, and the highest mean MSIS. scope: population: time series of the M1, M3 and tourism forecasting competitions | conditions: seasonal or non-seasonal CES selected by AICc C4 [descriptive; principal; support: empirical_quant E8 (E8.T1.r1, E8.T1.r2, E8.T1.r3, E8.T1.r4); limitations L3, L5; inherits C60, C62, C63, C64; verifiability open] On the [[D32:M4 competition data]], [[D17:CES]] has the lowest mean and median [[D15:MASE]] and [[D14:RMSSE]] and the lowest median [[D16:MSIS]] of [[D12:four methods, CES, ETS, ARIMA and the Theta method]], but not the lowest mean MSIS. scope: population: M4 competition time series | excludes: 10 series on which ETS broke down C5 [formal; principal; support: formal_proof E5; assumes C61; uses C16; verification unchecked; verifiability open] [[D1:CES]] has an underlying ARMA(2,2) model whose coefficients $\phi_1 = 2 - \alpha_0$, $\phi_2 = \alpha_0 + \alpha_1 - 2$, $\theta_1 = 2 - 2\alpha_0 + \alpha_1$ and $\theta_2 = 3\alpha_0 + \alpha_1 - 2 - \alpha_0^2 - \alpha_1^2$ are nonlinear functions of the same two parameters $\alpha_0$ and $\alpha_1$. C6 [formal; principal; support: formal_proof E5; uses C5; verification unchecked; verifiability open] The parameter space of the ARMA(2,2) model underlying [[D1:CES]] differs from that of the general ARMA(2,2) model. C7 [formal; principal; support: formal_proof E5; uses C5; verification unchecked; verifiability open] The parameter space of the ARMA(2,2) model underlying [[D1:CES]] is not restricted to stationary models. C8 [methodological; principal; support: argument W1; limitations L2; warrant W1; verifiability open] [[D4:seasonal CES]] models seasonality with lag-$m$ level and nonlinear trend states instead of seasonal indices. scope: data_type: seasonal univariate time series | conditions: additive error; seasonal period m C9 [methodological; principal; support: argument W1; uses C8; warrant W1; verifiability open] [[D4:seasonal CES]] has no seasonal indices that need renormalising. scope: data_type: seasonal univariate time series | conditions: additive error; seasonal period m C10 [methodological; principal; support: argument W1; uses C8, C31, C13, C14, C15; limitations L2; warrant W1; verifiability open] Model selection for [[D1:CES]] reduces to choosing between the seasonal and the non-seasonal model. scope: data_type: univariate time series, seasonal or non-seasonal | conditions: selection by an information criterion C11 [descriptive; principal; support: empirical_quant E2 (E2.T1.r5, E2.T1.r6); uses C35; limitations L3; verifiability open] On the [[D8:M1, M3 and tourism data]], [[D5:SCUM]] has lower mean and median [[D14:RMSSE]] and [[D15:MASE]] and lower median [[D16:MSIS]] than [[D6:SCUM(noCES)]], but higher mean MSIS. scope: population: time series of the M1, M3 and tourism forecasting competitions C12 [descriptive; principal; support: empirical_quant E8 (E8.T1.r5, E8.T1.r6); limitations L3, L5; verifiability open] On the [[D32:M4 data without 10 series]], [[D5:SCUM]] has lower mean and median [[D15:MASE]], [[D14:RMSSE]] and [[D16:MSIS]] than [[D6:SCUM(noCES)]]. scope: population: M4 competition time series | excludes: 10 series on which ETS broke down C16 [formal; supporting; support: formal_proof E1; assumes C60, C61; verification unchecked; verifiability open] [[D1:CES]] has the single-source-of-error state-space form $y_t = l_{t-1} + \varepsilon_t$, $l_t = l_{t-1} - (1 - \alpha_1) c_{t-1} + (\alpha_0 - \alpha_1) \varepsilon_t$, $c_t = l_{t-1} + (1 - \alpha_0) c_{t-1} + (\alpha_0 + \alpha_1) \varepsilon_t$, with level $l_t$ and [[D3:nonlinear trend component]] $c_t$. C17 [formal; supporting; support: formal_proof E3, cited B1, at p. 38, own earlier work; verification unchecked; verifiability open] [[D1:CES]] is stationary, with both eigenvalues of its transition matrix inside the unit circle, when $\alpha_1 < 5 - 2\alpha_0$, $\alpha_1 < 1$ and $\alpha_1 > 1 - \alpha_0$. C18 [formal; supporting; support: formal_proof E3; verification unchecked; verifiability open] When $\alpha_1 > 1$, both eigenvalues of the [[D1:CES]] transition matrix are real and one of them is greater than one. C19 [formal; supporting; support: formal_proof E3; uses C18; verification unchecked; verifiability open] When $\alpha_1 > 1$, [[D1:CES]] produces a non-stationary trajectory. C20 [formal; supporting; support: formal_proof E3; verification unchecked; verifiability open] When $\alpha_1 = 1$, [[D1:CES]] is equivalent to [[D19:ETS(A,N,N)]]. C21 [methodological; supporting; support: cited B14, own earlier work, cited B1, own earlier work; verifiability open] [[D1:CES]] is required to be stable, or at least forecastable. C22 [methodological; supporting; support: cited B14, own earlier work, cited B1, own earlier work; verifiability open] [[D1:CES]] is not required to be stationary. C23 [formal; supporting; support: formal_proof E4, cited B1, at pp. 172-173, own earlier work; verification unchecked; verifiability open] [[D1:CES]] is stable, with all eigenvalues of its discount matrix inside the unit circle, when $(\alpha_0 - 2.5)^2 + \alpha_1^2 > 1.25$, $(\alpha_0 - 0.5)^2 + (\alpha_1 - 1)^2 > 0.25$ and $(\alpha_0 - 1.5)^2 + (\alpha_1 - 0.5)^2 < 1.5$. C24 [formal; supporting; support: formal_proof E4; uses C17, C23; verification unchecked; verifiability open] The stable region of [[D1:CES]] contains both stationary and non-stationary values of the [[D2:complex smoothing parameter]]. C25 [formal; supporting; support: formal_proof E3; uses C26, C27, C28, C29, C30; verification unchecked; verifiability open] The [[D1:CES]] point forecast $w'F^{h-1}v_t$ follows one of five trajectories set by the [[D2:complex smoothing parameter]]: a flat line, exponential growth, exponential decline, harmonic oscillation converging to zero, or diverging harmonic oscillation. explains C36 explains C37 C26 [formal; supporting; support: formal_proof E3; verification unchecked; verifiability open] When $\alpha_1 = 1$, the [[D1:CES]] point forecast is a flat line. C27 [formal; supporting; support: formal_proof E3; verification unchecked; verifiability open] When $\alpha_1 > 1$, the [[D1:CES]] point forecast grows exponentially. C28 [formal; supporting; support: formal_proof E3; verification unchecked; verifiability open] When $(4 - \alpha_0^2)/4 < \alpha_1 < 1$, the [[D1:CES]] point forecast declines exponentially. C29 [formal; supporting; support: formal_proof E3; verification unchecked; verifiability open] When $1 - \alpha_0 < \alpha_1 < (4 - \alpha_0^2)/4$, the [[D1:CES]] point forecast is a harmonic oscillation converging to zero. C30 [formal; supporting; support: formal_proof E3; verification unchecked; verifiability open] When $0 < \alpha_1 < 1 - \alpha_0$, the [[D1:CES]] point forecast is a diverging harmonic oscillation. C32 [descriptive; supporting; support: empirical_quant E7 (E7.R1, E7.R2); inherits C60, C62, C63, C64; verifiability open] On M3 series 1664, which graphical analysis and [[D40:ADF test]] and [[D41:KPSS test]] indicate is stationary, [[D1:CES]] estimates the stationary [[D2:complex smoothing parameter]] $\alpha_0 + i\alpha_1 = 0.99999 + 0.99884i$. C33 [descriptive; supporting; support: empirical_quant E7 (E7.R3, E7.R4); inherits C60, C62, C63, C64; verifiability open] On the trended M3 series 2721, [[D1:CES]] estimates the non-stationary [[D2:complex smoothing parameter]] $\alpha_0 + i\alpha_1 = 1.48098 + 1.00346i$. C34 [descriptive; supporting; support: empirical_quant E2; inherits C60, C62, C63, C64; verifiability open] On the [[D8:M1, M3 and tourism data]], in [[D38:Nemenyi tests]] at the 5% level, [[D17:CES]] has the best mean rank of [[D12:four methods, CES, ETS, ARIMA and the Theta method]] on [[D14:RMSSE]] and on [[D15:MASE]]. scope: conditions: Nemenyi test on ranks of the error measures, significance level 0.05 C35 [descriptive; supporting; support: empirical_quant E2; verifiability open] On the [[D8:M1, M3 and tourism data]], in [[D38:Nemenyi test]]s at the 5% level [[D5:SCUM]] is statistically significantly better than every other method of [[D7:the six compared methods]], including [[D6:SCUM(noCES)]], on [[D14:RMSSE]], [[D15:MASE]] and [[D16:MSIS]]. scope: conditions: Nemenyi test on ranks of the error measures, significance level 0.05 C36 [descriptive; supporting; support: empirical_quant E9 (E9.T1); limitations L1; inherits C65, C60, C62, C63, C64; verifiability open] On the [[D33:3262 non-seasonal series]] of the [[D8:M1, M3 and tourism data]], [[D17:CES]] has lower mean and median [[D14:RMSSE]] and [[D15:MASE]] and lower median [[D16:MSIS]] than [[D19:ETS(A,N,N)]], but a higher mean MSIS. C37 [descriptive; supporting; support: empirical_quant E9 (E9.T1); inherits C65, C60, C62, C63, C64; verifiability open] On the [[D33:3262 non-seasonal series]] of the [[D8:M1, M3 and tourism data]], [[D17:CES]] has lower mean and median [[D14:RMSSE]] and [[D15:MASE]] and lower median [[D16:MSIS]] than [[D20:ETS(A,A,N)]], but a higher mean MSIS. C38 [descriptive; supporting; support: empirical_quant E9 (E9.T1); inherits C65, C60, C62, C63, C64; verifiability open] On the [[D33:3262 non-seasonal series]] of the [[D8:M1, M3 and tourism data]], [[D17:CES]] has a lower median [[D14:RMSSE]] than [[D21:ETS(A,Ad,N)]], but higher mean RMSSE, mean and median [[D15:MASE]], and mean and median [[D16:MSIS]]. C39 [descriptive; supporting; support: empirical_quant E9 (E9.T1); inherits C65, C60, C62, C63, C64; verifiability open] On the [[D33:3262 non-seasonal series]] of the [[D8:M1, M3 and tourism data]], [[D17:CES]] has lower mean and median [[D14:RMSSE]], [[D15:MASE]] and [[D16:MSIS]] than [[D22:ETS(M,M,N)]]. C40 [descriptive; supporting; support: empirical_quant E9 (E9.T1); inherits C65, C60, C62, C63, C64; verifiability open] On the [[D33:3262 non-seasonal series]] of the [[D8:M1, M3 and tourism data]], [[D17:CES]] has lower mean and median [[D14:RMSSE]] and [[D15:MASE]] and lower median [[D16:MSIS]] than [[D23:ETS(M,Md,N)]], but a higher mean MSIS. C60 [descriptive; supporting; assumption (statistical, unchecked); verifiability not_applicable] The error term $\varepsilon_t$ of [[D1:CES]] is additive white noise. C61 [descriptive; supporting; assumption (theoretical, untestable); verifiability not_applicable] There is no misspecification error, so the forecast error $e_t$ of [[D1:CES]] equals the model error $\varepsilon_t$. C62 [descriptive; supporting; assumption (statistical, unchecked); verifiability not_applicable] The error term $\varepsilon_t$ of [[D1:CES]] has no autocorrelation. C63 [descriptive; supporting; assumption (statistical, unchecked); verifiability not_applicable] The error term $\varepsilon_t$ of [[D1:CES]] has constant variance. C64 [descriptive; supporting; assumption (statistical, unchecked); verifiability not_applicable] The error term $\varepsilon_t$ of [[D1:CES]] is normally distributed. C65 [descriptive; supporting; assumption (methodological, unchecked); verifiability not_applicable] [[D9:ETS]] model selection with es() from the smooth package identifies correctly which series of the [[D8:M1, M3 and tourism data]] are seasonal. C41 [formal; secondary; support: formal_proof E1; uses C16; verification unchecked; verifiability open] The transition matrix of the [[D1:CES]] state-space form contains the smoothing parameters. C42 [methodological; secondary; support: cited B1, at p. 68, own earlier work; assumes C60, C64; verifiability open] The [[D39:Akaike information criterion]] of seasonal and non-seasonal [[D1:CES]] can be computed from a Gaussian likelihood of the same form as that of additive exponential smoothing models. C43 [formal; secondary; support: cited B1, at p. 96, own earlier work; assumes C60, C62, C63; uses C16; verifiability open] The conditional variance of [[D1:CES]] $h$ steps ahead, given the level and the [[D3:nonlinear trend component]], can be calculated from its state-space form in the same way as for pure additive ETS models. C44 [formal; secondary; support: formal_proof E3; uses C20; verification unchecked; verifiability open] When $\alpha_1$ is close to 1, [[D1:CES]] behaves like simple exponential smoothing with smoothing parameter $\alpha_0 - 1$. C45 [formal; secondary; support: formal_proof E3; uses C44; verification unchecked; verifiability open] When $\alpha_1$ is close to 1, the real part $\alpha_0$ of the [[D2:complex smoothing parameter]] of [[D1:CES]] is close to 1 for a stationary series, for which the smoothing parameter of simple exponential smoothing would be close to 0. C46 [formal; secondary; support: formal_proof E3; uses C44; verification unchecked; verifiability open] When $\alpha_1$ is close to 1, the real part $\alpha_0$ of the [[D2:complex smoothing parameter]] of [[D1:CES]] is close to 2 for a random walk. C47 [formal; secondary; support: formal_proof E6; uses C8; verification unchecked; verifiability open] [[D4:seasonal CES]] has an underlying SARIMA(2,0,2m+2)(2,0,0)$_m$ model. C53 [descriptive; secondary; support: empirical_quant E2; inherits C60, C62, C63, C64; verifiability open] On the [[D8:M1, M3 and tourism data]], in [[D38:Nemenyi tests]] at the 5% level, not all of the differences between [[D17:CES]] and [[D13:ETS, ARIMA and the Theta method]] on [[D14:RMSSE]] and [[D15:MASE]] are statistically significant. scope: conditions: Nemenyi test on ranks of the error measures, significance level 0.05 qualifies C3 C54 [descriptive; secondary; support: empirical_quant E9 (E9.T2); inherits C65, C60, C62, C63, C64; verifiability open] On the [[D34:2053 seasonal series]] of the [[D8:M1, M3 and tourism data]], [[D17:CES]] has lower mean and median [[D14:RMSSE]] and [[D15:MASE]] and lower median [[D16:MSIS]] than [[D27:ETS(A,N,A)]], but a higher mean MSIS. C55 [descriptive; secondary; support: empirical_quant E9 (E9.T2); inherits C65, C60, C62, C63, C64; verifiability open] On the [[D34:2053 seasonal series]] of the [[D8:M1, M3 and tourism data]], [[D17:CES]] has lower mean and median [[D14:RMSSE]] and [[D15:MASE]] and lower median [[D16:MSIS]] than [[D29:ETS(A,Ad,A)]], but a higher mean MSIS. C56 [descriptive; secondary; support: empirical_quant E9 (E9.T2); inherits C65, C60, C62, C63, C64; verifiability open] On the [[D34:2053 seasonal series]] of the [[D8:M1, M3 and tourism data]], [[D17:CES]] has lower mean and median [[D14:RMSSE]] and [[D15:MASE]] and lower median [[D16:MSIS]] than [[D30:ETS(M,M,M)]], but a higher mean MSIS. C57 [descriptive; secondary; support: empirical_quant E9 (E9.T2); inherits C65, C60, C62, C63, C64; verifiability open] On the [[D34:2053 seasonal series]] of the [[D8:M1, M3 and tourism data]], [[D17:CES]] has lower mean and median [[D14:RMSSE]] and [[D15:MASE]] and lower median [[D16:MSIS]] than [[D31:ETS(M,Md,M)]], but a higher mean MSIS. C58 [interpretive; secondary; support: empirical_quant E9 (E9.T2), argument W6; inherits C65; warrant W6; verifiability open] Information criteria can effectively identify when to switch between seasonal and non-seasonal [[D1:CES]]. C59 [interpretive; secondary; support: empirical_quant E2 (E2.T1), empirical_quant E8 (E8.T1), argument W7; inherits C60, C62, C63, C64; warrant W7; verifiability open] [[D17:CES]] is a robust forecasting approach. scope: excludes: forecasting scenarios for which CES is not appropriate, which the paper does not name CLAIMS WITHOUT EVIDENCE C13 [methodological; supporting; support: argument W2; uses C20, C26, C36; warrant W2; verifiability open] [[D1:CES]] encompasses level processes. C14 [methodological; supporting; support: argument W2; uses C27, C28, C39, C40; warrant W2; verifiability open] [[D1:CES]] encompasses multiplicative trend processes. C15 [methodological; supporting; support: argument W2; uses C37, C38; warrant W2; verifiability open] [[D1:CES]] approximates additive trend processes well. C31 [methodological; supporting; support: argument W3; uses C25; warrant W3; verifiability open] [[D4:seasonal CES]] can produce seasonality whose amplitude changes nonlinearly. C48 [formal; secondary; support: argument W4; uses C47; limitations L2; warrant W4; verification unchecked; verifiability open] [[D4:seasonal CES]] is a stationary or a non-stationary model depending on the values of its complex smoothing parameters. C49 [methodological; secondary; support: argument W3; uses C25; warrant W3; verifiability open] [[D4:seasonal CES]] can produce all the trajectories of non-seasonal [[D1:CES]]. C50 [methodological; secondary; support: argument W3; uses C25; warrant W3; verifiability open] [[D4:seasonal CES]] can model multiplicative or other shapes of seasonality when the level of the series does not change. C51 [methodological; secondary; support: argument W5; warrant W5; verifiability open] Non-seasonal [[D1:CES]] has 5 estimated parameters: two smoothing parameters, two initial states and the error variance. C52 [methodological; secondary; support: argument W5; warrant W5; verifiability open] [[D4:seasonal CES]] has $4 + 2m + 2 + 1$ estimated parameters: four smoothing parameters, $2m$ initial lagged values, two initial non-seasonal states and the error variance. EVIDENCE E1 [formal; foundation Complex numbers represented as real 2x2 matrices; linear algebra] Derivation of the state-space form of [[D1:CES]] from the complex-valued method, representing complex variables as real vectors and matrices. assumes C60, C61 derivation: in the source E2 [quant; design benchmark; data_origin secondary; n 5315] [[D17:CES with automatic seasonal selection]] compared with [[D9:ETS]], [[D10:ARIMA]], the [[D11:Theta method]], [[D5:SCUM]] and [[D6:SCUM(noCES)]] on the [[D8:M1, M3 and tourism data]], by error measures on a single holdout per series, with Nemenyi tests of the differences. setup: studied D17, D9, D10, D11, D5, D6; cases D8; procedure D36, D38; lens D14, D15, D16 E2.R1 D16 = 523.034 {D7: D17, series: M1 series 47} E2.R2 D16 = 25.59 {D7: D9, series: M1 series 47, ETS model selected: ETS(M,N,N)} E2.T1 Mean and median RMSSE, MASE and MSIS of six methods over the 5,315 series of the M1, M3 and tourism competitions | row | D7 | D14 (mean) | D15 (mean) | D16 (mean) | D14 (median) | D15 (median) | D16 (median) | | r1 | D17 | 1.959 | 2.272 | 3.465 | 1.17 | 1.298 | 0.869 | | r2 | D9 | 1.97 | 2.263 | 2.258 | 1.181 | 1.323 | 0.925 | | r3 | D10 | 2.134 | 2.482 | 3.335 | 1.271 | 1.419 | 0.988 | | r4 | D11 | 1.965 | 2.252 | 2.531 | 1.238 | 1.377 | 0.895 | | r5 | D5 | 1.867 | 2.146 | 2.333 | 1.128 | 1.256 | 0.844 | | r6 | D6 | 1.911 | 2.191 | 2.223 | 1.143 | 1.279 | 0.859 | E3 [formal; foundation Linear algebra: eigenvalues of the transition matrix of a linear state-space model] Analysis of the eigenvalues of the [[D1:CES]] transition matrix: the stationarity condition, the non-stationary case, the special case equivalent to [[D19:ETS(A,N,N)]], the five forecast trajectories and the connection with simple exponential smoothing. assumes C61 derivation: in the source E4 [formal; foundation Invertibility conditions of ARMA(2,2) models; algebra of inequalities] Derivation of the stability region of [[D1:CES]] from the invertibility conditions of its underlying ARMA(2,2) model. assumes C61 derivation: in the source E5 [formal; foundation Algebra of backshift-operator polynomials] Derivation of the ARMA(2,2) model underlying [[D1:CES]] from its state-space form, by eliminating the [[D3:nonlinear trend component]]. assumes C61 derivation: in the source E6 [formal; foundation Algebra of backshift-operator polynomials and matrix inversion] Derivation of the SARIMA model underlying [[D4:seasonal CES]] from its state-space form. assumes C61 derivation: in the source E7 [quant; design other; data_origin secondary; n 2] [[D1:CES]] fitted with ces() from smooth to [[D42:M3 series 1664 and 2721]], each split into in-sample and holdout parts; stationarity of each series assessed by graphical analysis and [[D40:ADF test]] and [[D41:KPSS test]]; point forecasts and 95% prediction intervals. setup: studied D1; cases D42; procedure D40, D41; lens D2 E7.R1 D2 = 0.99999 {part: real part alpha_0, series: M3 series 1664} E7.R2 D2 = 0.99884 {part: imaginary part alpha_1, series: M3 series 1664} E7.R3 D2 = 1.48098 {part: real part alpha_0, series: M3 series 2721} E7.R4 D2 = 1.00346 {part: imaginary part alpha_1, series: M3 series 2721} E8 [quant; design benchmark; data_origin secondary; n 99990] The same six methods compared on the [[D32:M4 data without 10 series]] by error measures on a single holdout per series, without a significance test. setup: studied D17, D9, D10, D11, D5, D6; cases D32; procedure D37; lens D15, D14, D16 E8.T1 Mean and median MASE, RMSSE and MSIS of six methods on the M4 competition data, excluding 10 series | row | D7 | D15 (mean) | D14 (mean) | D16 (mean) | D15 (median) | D14 (median) | D16 (median) | | r1 | D17 | 2.921 | 2.339 | 2.379 | 1.803 | 1.529 | 0.592 | | r2 | D9 | 3.31 | 2.688 | 4.751 | 1.831 | 1.542 | 0.632 | | r3 | D10 | 3.037 | 2.436 | 2.272 | 1.903 | 1.606 | 0.617 | | r4 | D11 | 2.98 | 2.394 | 2.13 | 1.865 | 1.574 | 0.597 | | r5 | D5 | 2.737 | 2.198 | 1.751 | 1.742 | 1.475 | 0.568 | | r6 | D6 | 2.784 | 2.241 | 1.798 | 1.769 | 1.498 | 0.583 | E9 [quant; design benchmark; data_origin secondary; n 5315] [[D17:CES with automatic seasonal selection]] compared with the [[D18:non-seasonal special cases of ETS]] on the [[D33:non-seasonal subset]] and with the [[D24:seasonal special cases of ETS]] on the [[D34:seasonal subset]], by error measures on a single holdout per series. setup: studied D17, D18, D24; cases D33, D34; procedure D36; lens D14, D15, D16 derived from E2 (subset) E9.T1 Mean and median RMSSE, MASE and MSIS of CES and five non-seasonal ETS models on the non-seasonal subset of the M1, M3 and tourism series {D35: D33} | row | D25 | D14 (mean) | D15 (mean) | D16 (mean) | D14 (median) | D15 (median) | D16 (median) | | r1 | D17 | 2.534 | 2.949 | 4.269 | 1.654 | 1.88 | 0.863 | | r2 | D19 | 2.823 | 3.234 | 3.104 | 1.95 | 2.228 | 0.917 | | r3 | D20 | 2.67 | 3.095 | 3.785 | 1.694 | 1.919 | 0.907 | | r4 | D21 | 2.502 | 2.907 | 3.769 | 1.66 | 1.863 | 0.854 | | r5 | D22 | 3.185 | 3.698 | 4.768 | 1.875 | 2.09 | 0.96 | | r6 | D23 | 2.691 | 3.16 | 3.546 | 1.773 | 1.975 | 0.971 | E9.T2 Mean and median RMSSE, MASE and MSIS of CES and five seasonal ETS models on the seasonal subset of the M1, M3 and tourism series {D35: D34} | row | D26 | D14 (mean) | D15 (mean) | D16 (mean) | D14 (median) | D15 (median) | D16 (median) | | r1 | D17 | 1.044 | 1.196 | 2.188 | 0.756 | 0.844 | 0.877 | | r2 | D27 | 1.142 | 1.297 | 2.037 | 0.803 | 0.894 | 0.92 | | r3 | D28 | 1.135 | 1.295 | 2.247 | 0.802 | 0.877 | 0.934 | | r4 | D29 | 1.094 | 1.245 | 2.155 | 0.797 | 0.876 | 0.946 | | r5 | D30 | 1.171 | 1.352 | 2.153 | 0.806 | 0.875 | 1.057 | | r6 | D31 | 1.103 | 1.26 | 1.934 | 0.789 | 0.875 | 1.006 | WARRANTS W1 [deductive; premises C8, C31, C13, C14, C15] The seasonal model takes the level and nonlinear trend states at lag m instead of carrying a set of seasonal components, so there are no seasonal indices to renormalise. Since non-seasonal CES captures both level and trend series without a choice of trend form, the only choice left is whether to include the lag-m states. W2 [deductive; premises C20, C26, C27, C28] With the imaginary part of the smoothing parameter equal to one CES is the local level model ETS(A,N,N); with it above one the forecasts grow exponentially, as in multiplicative trend models; and the flexibility of its trajectories lets it follow trends that the level and trend classification of exponential smoothing captures only with separate models. W3 [deductive; premises C25] Seasonal CES consists of two CES parts, one at lag 1 and one at lag m, and each part follows one of the CES forecast trajectories, so the seasonal part can grow, decline or oscillate in its own way. The non-seasonal level can become negative while the lagged level stays positive, and the two levels evolve independently, so the seasonal part can change its amplitude and shape without any change in the level of the series. W4 [analogical; premises C47, C5] Non-seasonal CES has an underlying ARMA(2,2) model that is stationary or not depending on its complex smoothing parameter; by analogy, the authors state the same of the SARIMA model underlying seasonal CES. No stationarity condition is derived for seasonal CES. W5 [deductive] The count follows from the model equations: each complex smoothing parameter has two real parts, each pair of level and nonlinear trend states needs initial values (m of each for the lag-m pair), and the error variance is estimated. W6 [inductive] On the seasonal subset of the M1, M3 and tourism data, CES with its seasonal or non-seasonal form chosen by AICc has lower RMSSE and MASE than every seasonal ETS model, which the authors take to show that the criterion chooses the seasonal form when it is needed. This is an argument from the accuracy results only: no experiment compares the selection by information criteria with other ways of choosing between seasonal and non-seasonal CES. W7 [inductive] CES has lower RMSSE than ETS, ARIMA and the Theta method on the M1, M3, tourism and M4 data and lowers the errors of the combination it is included in, which the authors take as evidence that it is robust, while acknowledging that it does not suit every forecasting scenario. Robustness itself is not defined or tested: no experiment examines how the accuracy of CES varies across conditions or departures from its assumptions. LIMITATIONS L1 [model_specification; affects C3, C36, C37, C38, C39, C40; could_bias_magnitude; mitigation none] CES produces very large errors on a few series: on M1 series 47 it fails to identify the trend and its MSIS is 523.034 against 25.590 for ETS(M,N,N), which inflates its mean MSIS, and on the non-seasonal subset it fails on two series. L2 [model_specification; affects C8, C10, C47; reduces_precision; mitigation none] The parameter restrictions under which seasonal CES is stable are not derived. L3 [sample_size; affects C3, C4, C11, C12; reduces_transferability; mitigation none] The effect of the length of a time series on the accuracy of CES is not analysed. L4 [model_specification; affects C16; reduces_transferability; mitigation none] CES is formulated only with an additive error term; a version with a multiplicative error term is not developed. L5 [sampling; affects C4, C12; could_bias_magnitude; mitigation none] The 10 M4 series on which ETS broke down were excluded from the M4 comparison, so the series were selected on the failure of one of the compared methods. OPEN QUESTIONS O1 [from C11] Whether CES improves the SCUM combination because it captures long-term relations better than ETS, ARIMA and the Theta method, and nonlinear trends without separating level and trend, is not established. ARTEFACTS A1 [data] M1 competition data A2 [data] M3 competition data A3 [data] Tourism forecasting competition data A4 [software] smooth package for R (ces, auto.ces, es, auto.ssarima) A5 [software] forecast package for R (thetaf) A6 [software] greybox package for R (rmcb, Nemenyi test) A7 [data] M4 competition data A8 [software] R, the environment for statistical computing in which the experiments were run A9 [manuscript] The published article (open access, CC BY). REFERENCES B1 10.1007/978-3-540-71918-2 (doi) B2 10.1016/S0169-2070(01)00110-8 (doi) B3 10.1016/j.ijforecast.2013.09.006 (doi) B4 10.1016/j.ijforecast.2010.04.006 (doi) B5 10.1016/j.eswa.2020.113637 (doi) B6 10.1002/for.3980010202 (doi) B7 10.1016/S0169-2070(00)00057-1 (doi) B8 10.1016/j.ijforecast.2010.04.009 (doi) B9 10.1016/j.ijforecast.2019.04.014 (doi) B10 https://github.com/config-i1/smooth (url) B11 10.18637/jss.v027.i03 (doi) B12 https://github.com/config-i1/greybox (url) B13 10.1016/j.ijforecast.2019.01.006 (doi) B14 10.1080/01621459.1997.10473684 (doi) B15 10.1080/00207543.2019.1600764 (doi) B16 10.1016/S0169-2070(00)00066-2 (doi) B17 10.1016/j.ijforecast.2021.11.013 (doi) B18 10.1016/j.ijforecast.2006.03.001 (doi) B19 10.1198/016214506000001437 (doi) B20 Demšar, J. (2006). Statistical comparisons of classifiers over multiple data sets. The Journal of Machine Learning Research, 7, 1–30. B21 10.1109/TAC.1974.1100705 (doi) B22 10.1080/01621459.1979.10482531 (doi) B23 10.1016/0304-4076(92)90104-Y (doi) B24 R Core Team (2021). R: A language and environment for statistical computing. R Foundation for Statistical Computing. https://www.R-project.org/ B25 10.1002/nav.22074 (doi) SCORES Scores are computed by rules from the claimset file (EntryLib rubric 0.7): they check the record, not the science, and may change as the rubric and reviews develop. Support of the paper: 67 (medium), the mean of its 12 principal claims; the lowest is C10 at 43. C1: Support 65 (medium); novelty: extends earlier work, by the authors C2: Support 70 (medium); novelty: new, by the authors C3: Support 73 (medium); novelty: new, by the authors C4: Support 83 (high); novelty: new, by the authors C5: Support 66 (medium); novelty: new, by the authors C6: Support 66 (medium); novelty: new, by the authors C7: Support 66 (medium); novelty: new, by the authors C8: Support 57 (low); novelty: new, by the authors C9: Support 57 (low); novelty: new, by the authors C10: Support 43 (low); novelty: new, by the authors C11: Support 79 (medium); novelty: new, by the authors C12: Support 83 (high); novelty: new, by the authors C13: Support 59 (low); novelty: new, by the authors C14: Support 59 (low); novelty: new, by the authors C15: Support 74 (medium); novelty: new, by the authors C16: Support 71 (medium); novelty: new, by the authors C17: Support 80 (high); novelty: new, by the authors C18: Support 75 (medium); novelty: new, by the authors C19: Support 75 (medium); novelty: new, by the authors C20: Support 75 (medium); novelty: new, by the authors C21: Support 70 (medium); novelty: new, by the authors C22: Support 70 (medium); novelty: new, by the authors C23: Support 80 (high); novelty: new, by the authors C24: Support 75 (medium); novelty: new, by the authors C25: Support 75 (medium); novelty: new, by the authors C26: Support 75 (medium); novelty: new, by the authors C27: Support 75 (medium); novelty: new, by the authors C28: Support 75 (medium); novelty: new, by the authors C29: Support 75 (medium); novelty: new, by the authors C30: Support 75 (medium); novelty: new, by the authors C31: Support 65 (medium); novelty: new, by the authors C32: Support 90 (high); novelty: new, by the authors C33: Support 90 (high); novelty: new, by the authors C34: Support 90 (high); novelty: new, by the authors C35: Support 90 (high); novelty: new, by the authors C36: Support 74 (medium); novelty: new, by the authors C37: Support 74 (medium); novelty: new, by the authors C38: Support 74 (medium); novelty: new, by the authors C39: Support 74 (medium); novelty: new, by the authors C40: Support 74 (medium); novelty: new, by the authors C41: Support 71 (medium); novelty: new, by the authors C42: Support 66 (medium); novelty: new, by the authors C43: Support 69 (medium); novelty: new, by the authors C44: Support 75 (medium); novelty: new, by the authors C45: Support 75 (medium); novelty: new, by the authors C46: Support 75 (medium); novelty: new, by the authors C47: Support 57 (low); novelty: new, by the authors C48: Support 57 (low); novelty: new, by the authors C49: Support 65 (medium); novelty: new, by the authors C50: Support 65 (medium); novelty: new, by the authors C51: Support 60 (medium); novelty: new, by the authors C52: Support 60 (medium); novelty: new, by the authors C53: Support 90 (high); novelty: new, by the authors C54: Support 80 (high); novelty: new, by the authors C55: Support 80 (high); novelty: new, by the authors C56: Support 80 (high); novelty: new, by the authors C57: Support 80 (high); novelty: new, by the authors C58: Support 80 (high); novelty: new, by the authors C59: Support 90 (high); novelty: new, by the authors Open problems in the principal claims and what they rest on: P1 [support; up to -30; C1, C2, C3, C5, C6, C7, C8, C9, C10, C11, C13, C14, C15, C16, C17, C18, C19, C20, C21, C22, C23, C24, C25, C26, C27, C28, C29, C30, C31, C34, C35, C36, C37, C38, C39, C40, C41, C42, C43, C44, C45, C46, C47, C48, C49, C50, C51, C52, C53, C54, C55, C56, C57, C58, C59] Rests on argument only; an unchecked proof; evidence named without the observations the claim reports; an argument from other claims; cited work only. To remove it: Support the claim with evidence the claimset records, citing the result entries or table rows it reports, or have its proof checked; for cited work, read the work in full and quote the passage the claim rests on from a captured copy (claimset extension review). P2 [context; up to -10; C1, C10, C13, C14, C15, C21, C22, C31, C32, C33, C36, C37, C38, C39, C40, C42, C49, C50, C51, C52, C54, C55, C56, C57, C58] No scope: where the claim holds is not stated. To remove it: Give the claim a scope: population, setting, period, data, conditions, exclusions. P3 [open; up to -6; C3, C10, C13, C14, C15, C36, C37, C38, C39, C40] Limitation L1 (could bias magnitude, not mitigated): CES produces very large errors on a few series: on M1 series 47 it fails to identify the trend and its MSIS is 523.034 against 25.590 for ETS(M,N,N), which inflates its mean MSIS, and on the non-seasonal subset it fails on two series. To remove it: Mitigate it with evidence, or narrow the claim's scope so that it no longer applies. P4 [plan; up to -5; C3, C4, C8, C9, C10, C11, C12, C47, C48] A principal claim reported without a plan recorded beforehand. To remove it: Plan the measurement in the lab record or a registration before running it, or publish the lab record that shows every run. P5 [context; up to -5; C2, C3, C4, C5, C6, C7, C8, C9, C10, C11, C12, C47, C48] A principal claim related to no earlier work. To remove it: Relate the claim to the earlier work it extends, addresses or departs from. P6 [open; up to -6; C4, C12] Limitation L5 (could bias magnitude, not mitigated): The 10 M4 series on which ETS broke down were excluded from the M4 comparison, so the series were selected on the failure of one of the compared methods. To remove it: Mitigate it with evidence, or narrow the claim's scope so that it no longer applies. P7 [open; up to -3; C5, C6, C7, C16, C41, C42, C43] Assumption C60 is not checked: The error term $\varepsilon_t$ of [[D1:CES]] is additive white noise. To remove it: Test the assumption, or analyse how sensitive the claim is to it. P8 [open; up to -3; C8, C9, C10, C47, C48] Limitation L2 (reduces precision, not mitigated): The parameter restrictions under which seasonal CES is stable are not derived. To remove it: Mitigate it with evidence, or narrow the claim's scope so that it no longer applies. P9 [open; up to -1; C3, C4, C11, C12] Limitation L3 (reduces transferability, not mitigated): The effect of the length of a time series on the accuracy of CES is not analysed. To remove it: Mitigate it with evidence, or narrow the claim's scope so that it no longer applies. P10 [open; up to -1; C5, C6, C7, C16, C41, C43] Limitation L4 (reduces transferability, not mitigated): CES is formulated only with an additive error term; a version with a multiplicative error term is not developed. To remove it: Mitigate it with evidence, or narrow the claim's scope so that it no longer applies. HOW TO CITE This claimset: Svetunkov, I. (guarantor) (2026). Complex exponential smoothing. EntryLib. entrylib:2026.000012@v1. https://entrylib.org/p/2026.000012@v1 One claim of it: by its identifier, as entrylib:2026.000012@v1#C1 for claim C1. Each principal claim has its own page, as https://entrylib.org/p/2026.000012/C1.