That is, Ypi is an unbiased predictor of E(Yi) = Xiβ, with heterogeneous variance Var(Ypi) = s2Xi(X'X)-1Xi.
If Xi is known for i=N+1,N+2,..., then Ypi = Xib is called the unconditional forecast of Yi. If Xi is not known and must be forecasted first, then Ypi = Xib is called the conditional forecast of Yi.
If Yi is known for i=N+1,N+2,..., then Ypi = Xib is called the ex-post forecast of Yi. If Yi is not known, then Ypi = Xib is called the ex-ante forecast of Yi.
For ex-post forecasts Ypi, i=N+1,N+2,..., the difference Yi-Ypi is called the forecast error. That is:
ei | = Yi-Ypi (i=N+1,N+2,...) |
= [Yi-E(Yi)] + [E(Yi)-Ypi] | |
= [Yi-E(Yi)] + [E(Ypi)-Ypi] | |
= εi + Xi(β-b) | |
(1) (2) |
E(ei) | = 0 |
Var(ei) | = Var(εi) + Var(Ypi) |
= σ2[1+Xi(X'X)-1Xi] |
By normality assumption ε|X ~ Normal(0,σ2I), we have ei ~ Normal(0,σ2i), where σ2i = σ2[1+Xi(X'X)-1Xi], for i=N+1,N+2,...
Therefore, ei/σi = (Yi-Ypi)/(σ2[1+Xi(X'X)-1Xi])½ ~ Normal(0,1).
Subsituting the unknown σ2 by the unbiased sample estimator s2, we have
ei/si =
(Yi-Ypi)/(s2[1+Xi(X'X)-1Xi])½
~ t(N-K).
Where
si = (s2[1+Xi(X'X)-1Xi])½
Ypi-tα/2si = lower bound of forecast
Ypi+tα/2si = upper bound of forecast
ln(Y) = Xβ + ε
The normality assumption of ε implies that Y is log-normal.
Let Z = ln(Y), or Y = eZ = exp(Z). Then:
E(Y) = E(exp(Z)) =
exp(Xβ+σ2/2)
Var(Y) = Var(exp(Z)) =
exp(2Xβ+σ2)
[exp(σ2)-1]
As N → ∞, Y ~ Normal(E(Y),Var(Y)).
Note: Median(Y) = exp(Xβ) < E(Y).
Pr[ln(Ypi)-tα/2si ≤ ln(Yi) ≤ ln(Ypi)+tα/2si] = 1-α
Ypi = exp(ln(Ypi)) = exp(Xib)
Ypi is the unbiased predictor of the Median of Yi or exp(Xib), which always under-predict the Mean of Yi or E(Yi) = exp(Xib+si2/2). Therefore, the unbiased predictor of E(Yi) should be:
Ypi = exp(Xib+si2/2), with the variance
Var(Ypi) = exp(2Xib+si2)[exp(si2)-1]
As N → ∞, Ypi ~ Normal(E(Ypi),Var(Ypi)).
Given a level of significance α > 0, the confidence interval of forecast of Yi is defined by:
Pr[Ypi-zα/2si ≤ Yi ≤ Ypi+zα/2si] = 1-α
where zα/2 is the critical value of standard normal and si = Var(Ypi)½.
Evaluating the forecast performance by comparing the actuals and the predicted:
YN+1,YN+2,...,YN+F |
↓ |
YpN+1,YpN+2,...,YpN+F |
Let
Ym = ∑i=N+1,...,N+FYi/F |
Ypm = ∑i=N+1,...,N+FYpi/F |
σ2Y = ∑i=N+1,...,N+F(Yi-Ym)2/F |
σ2Yp = ∑i=N+1,...,N+F(Ypi-Ypm)2/F |
σY,Yp = ∑i=N+1,...,N+F(Yi-Ym)(Ypi-Ypm)/F |
⌈ | ∑i=N+1,...,N+F(Yi-Ym)(Ypi-Ypm) | ⌉2 | |
r2 = | | | | | |
⌊ | [∑i=N+1,...,N+F(Yi-Ym)2]½ [∑i=N+1,...,N+F(Ypi-Ypm)2]½ | ⌋ |
That is, r2 = (σY,Yp/σYσYp)2
MSE | = ∑i=N+1,...,N+F(Yi-Ypi)2/F |
= ∑i=N+1,...,N+F[(Yi-Ym)+(Ym-Ypm)-(Ypi-Ypm)]2/F | |
= (Ym-Ypm)2 +
∑i=N+1,...,N+F(Yi-Ym)2/F + ∑i=N+1,...,N+F(Ypi-Ypm)2/F - 2∑i=N+1,...,N+F(Yi-Ym)(Ypi-Ypm)/F | |
= (Ym-Ypm)2 + σ2Y + σ2Yp - 2rσYσYp | |
= (Ym-Ypm)2 + (σY-σYp)2 + 2(1-r)σYσYp | |
= Bias Component + Variance Component + Covariance Component | |
= (Ym-Ypm)2 + (σYp-rσY)2 + (1-r2)σ2Y | |
= Bias Component + Regression Component + Disturbance Component |
Then, UM+US+UC = 1 and UM+UR+UD = 1. Idealy UM → 0 and US and UR are small for a good forecast.
(∑i=N+1,...,N+F(Yi-Ypi)2/F)½ | |
U = | |
(∑i=N+1,...,N+FYi2/F)½ + (∑i=N+1,...,N+FYpi2/F)½ | |
RMSE | |
= | |
(∑i=N+1,...,N+FYi2/F)½ + (∑i=N+1,...,N+FYpi2/F)½ |
Dik = | 1 if observation i is in the season k, k=1,2,...,K |
0 otherwise |