跳到主要内容

Macro by Mark 的 Cookie.我们使用必要 Cookie 处理登录、语言和安全。经您同意,我们还会将可选 Cookie 用于保存偏好、分析,以及通过 Meta(Facebook)Pixel 衡量广告效果。请阅读我们的Cookie 政策。

Macro by MarkMacro by Mark
  • 首页
加载中…
加载中…
← 模型
Macro概念政策模型

ARIMA Model

Autoregressive integrated moving average - the workhorse univariate forecaster for stationary or differenced macro series.

概述图表数学参考资料

Graph workspace

Time-series forecast cone

Graph workspace

The canonical visualization for ARIMA is a time-series plot with the historical series, a fitted in-sample path, and an out-of-sample forecast cone showing point forecasts with expanding confidence bands. This directly communicates the three things a practitioner cares about: how well the model tracks history, where the forecast is going, and how uncertain that forecast is at each horizon.

Chart axis metadata
xTime (period index)integer period · t=1t = 1t=1 to T+HT + HT+H, where T≈80T \approx 80T≈80 and H≈12H \approx 12H≈12ySeries levelindex units · Starts at 100. With d=1d = 1d=1 the level is a random walk plus drift, so it wanders rather than reverting: the plotted band widens by roughly 80c/(1−ϕ1)80 c / (1 - \phi_1)80c/(1−ϕ1​) over the sample instead of staying inside a fixed 80-120 range.

Shaded ranges come from assumed parameters in a simulated example. They are not estimated from observed data.

Time-series forecast with illustrative ranges: Series level (index units) by Time (period index) static chart preview showing History, Forecast
xTime (period index)integer period · t=1t = 1t=1 to T+HT + HT+H, where T≈80T \approx 80T≈80 and H≈12H \approx 12H≈12ySeries levelindex units · Starts at 100. With d=1d = 1d=1 the level is a random walk plus drift, so it wanders rather than reverting: the plotted band widens by roughly 80c/(1−ϕ1)80 c / (1 - \phi_1)80c/(1−ϕ1​) over the sample instead of staying inside a fixed 80-120 range.

Current results

1-step forecast

Explain: 1-step forecast
y^T+1=yT+c+ϕ1wT+θ1εT\hat{y}_{T+1} = y_T + c + \phi_1 w_T + \theta_1 \varepsilon_Ty^​T+1​=yT​+c+ϕ1​wT​+θ1​εT​

The model's best guess for next period, combining drift, AR persistence, and the last MA correction.

165.08

index level

12-step forecast

Explain: 12-step forecast
y^T+12\hat{y}_{T+12}y^​T+12​

Where the model expects the series to be in 12 periods. For large ϕ1\phi_1ϕ1​, this differs from the drift-only extrapolation.

173.01

index level

95% interval width at h=12

Explain: 95% interval width at h=12
2×1.96×σ∑j=011(∑k=0jψk)22 \times 1.96 \times \sigma \sqrt{\sum_{j=0}^{11}\left(\sum_{k=0}^{j}\psi_k\right)^2}2×1.96×σ∑j=011​(∑k=0j​ψk​)2​

Level-series uncertainty after cumulating the ARMA innovation weights through the first difference.

48.35

index units

In-sample RMSE

Explain: In-sample RMSE
1T∑t=1T(yt−y^t)2\sqrt{\frac{1}{T} \sum_{t=1}^{T} (y_t - \hat{y}_t)^2}T1​∑t=1T​(yt​−y^​t​)2​

Average magnitude of one-step-ahead in-sample prediction errors. A rough gauge of model fit.

0.9409

index units

Controls

Shocks fade steadily
Explain: How long growth shocks last

AR(1) coefficient: 0.7

How slowly a shock to the growth rate dies away; higher values keep it alive longer.

Increasing ϕ1\phi_1ϕ1​ makes the forecast path curve more slowly toward the long-run drift, and widens the forecast cone because MA(∞)\text{MA}(\infty)MA(∞) coefficients decay slower.

0.3
Explain: Weight on last period's surprise

How much last period's forecast error carries into the next forecast.

Changing θ1\theta_1θ1​ shifts the first forecast step relative to the last observation and alters how quickly the forecast cone expands in the first few periods.

1
Explain: Typical size of forecast errors

How big one-step forecast errors typically are; larger values widen the forecast bands.

Scaling σ\sigmaσ scales every forecast interval in proportion. It moves the point forecast too, since the history the forecast starts from is drawn with the same scaled shocks: at the other defaults the 1-step forecast is about 146 at σ=0.5\sigma = 0.5σ=0.5, 165 at 1 and 204 at 2.

0.1
Explain: Drift

Constant in the ARMA equation for the differenced series. For d=1d=1d=1, the long-run average growth rate per period is c/ϕ(1)=c/(1−ϕ1)c / \phi(1) = c / (1 - \phi_1)c/ϕ(1)=c/(1−ϕ1​), not ccc itself: at the defaults c=0.1c = 0.1c=0.1 and ϕ1=0.7\phi_1 = 0.7ϕ1​=0.7 this is 0.1 / 0.3, about 0.33.

Changing drift tilts the forecast path, and the history drawn with it, up or down. Over the 12 forecast steps the last observed growth, fading at rate ϕ1\phi_1ϕ1​, adds to the drift: at the other defaults c=−0.1c = -0.1c=−0.1 still ends the forecast about 0.6 above its first step, and only from about c=−0.15c = -0.15c=−0.15 down does it trend downward.

Keep going

Read next

Continue with a related concept, model, or example.

Empirical model

ARDL / Bounds Testing

a contrasting take on ARIMA

Contrast it →
Empirical model

Bridge Equations

a contrasting take on ARIMA

Contrast it →

继续学习

概念

  • Real and nominal→
  • Business cycles→
  • Prices and inflation→

政策

  • Monetary policy→
  • Fiscal policy→

模型

  • SARIMA→
  • ETS→
  • VAR→
Macro by Mark
帮助
隐私政策使用条款伦理与合规

© 2026 Mark Jayson Nation LLC