Methods and Motivation

Building a Forecast

Given series \(\mathcal{Y}_T:=\{Y_t\}_{t=1}^T\)
Forecast horizon \(h\)
- Number of periods out we want to forecast
Goal: predict the value of \(y_{T+h}\)
A forecasting method is a function that takes as input data \(\mathcal{Y}_T\) and possibly other inputs, and produces a forecast \[\widehat{y}_{t+h}=f(\mathcal{Y}_T,\text{other stuff})\]
Today
- Overview of some example methods
- How to evaluate results

Simple Forecast Methods

Trivial: always predict 0, regardless of the data \[f(\mathcal{Y}_T)=0\]
- Simple, easy to compute, stable
Naïve: Just pick the last value \[f(\mathcal{Y}_T)=y_T\]
- The next period will be the same as the current period
- Simple and straightforward, and uses the data
- Also called “random walk forecast”
Mean: Take the average past value \[f(\mathcal{Y}_T)=\frac{1}{T}\sum_{t=1}^T y_t\]
- Still fairly simple, now uses all the data

Implementation

Apply methods to a monthly series of a Manufacturer’s inventories of Condensed Milk, 1971-1980
Typical company forecasting task: predict inventories to avoid over-accumulation or stock-outs

library(fpp2)
data(condmilk) #Inventory Data
#Forecast using each of above methods
milknaive<-naive(condmilk,h=12) #Naive forecast, 12 months out
milkmean<-meanf(condmilk,h=12) #Mean forecast, 12 months out
trivialFC<-c(0,0,0,0,0,0,0,0,0,0,0,0)
milktrivial<-ts(trivialFC,start=c(1981,1),frequency=12) #Trivial forecast, 12 months out

#Plot the forecasts
autoplot(condmilk,series="Observed")+
  autolayer(milknaive, PI=FALSE, series="Naive")+
  autolayer(milktrivial, series="Trivial")+
  autolayer(milkmean, PI=FALSE,series="Mean")+
  ggtitle("Forecasts for Monthly Milk Stocks") +
  guides(colour=guide_legend(title="Series and Forecasts"))

Plot

Methods with parameters

Going beyond the simplest methods, one can construct families of methods which are similar
- Differ by one or more additional inputs to the method, which must be chosen
- Extra inputs are called parameters
- Once parameters are set, method again depends only on the data
Denote parameter by \(\theta\), and set of possible parameters as \(\Theta\)
A parametric forecast method is a set of methods

\[\{f(\mathcal{Y}_T,\theta,\text{other stuff})\}_{\theta\in\Theta}\]

Examples of Parametric Methods

Trivial: always predict the same number, \(\theta\), regardless of the data \[f(\mathcal{Y}_T,\theta)=\theta\]
- \(\theta=0\) gives the trivial forecast
Autoregression: Take \(\theta\) times the last value \[f(\mathcal{Y}_T,\theta)=\theta y_T\]
- \(\theta=1\) gives the naïve forecast
- Extend to horizon \(h\) by \(\widehat{y}_{T+h}=\theta^h y_T\)
Weighted Average: Take average of past value + a constant, with different weights on different times \[f(\mathcal{Y}_T,\{\theta_t\}_{t=1}^{T})=\theta_0+\sum_{t=1}^T \theta_t y_t\]
- Extremely flexible, but requires many choices
  - Picking \(\theta_t=\frac{1}{T}\) for all t,\(\theta_0=0\) gives unweighted average
  - Picking \(\theta_T=\theta\), \(\theta_t=0\) for all \(t<T\) gives autoregression

Structured parametric models

Arbitrary weighted averages can be too complicated, but some flexibility can help: use special cases
Autoregression of order p, abbreviated as AR(p)
- Weighted sum of \(p\) most recent values (and a constant) \[f(\mathcal{Y}_T,\{\theta_j\}_{j=0}^{p})=\theta_0+\sum_{j=0}^{p-1} \theta_{j+1} y_{T-j}\]
Simple Exponential Smoothing
- Weighted average where weights decline by fixed amount each time period
- Use all data, with more weight on recent past
- Takes two parameters \(\alpha\in[0,1]\) and \(\ell_0\) \[f(\mathcal{Y}_T,\alpha,\ell_0)=(1-\alpha)^T\ell_0+\sum_{j=0}^{T-1} \alpha(1-\alpha)^jy_{T-j}\]
- Higher \(\alpha\) gives more weight to recent data, up to \(\alpha=1\) naïve forecast

Choosing parameters

Applying a parametric method requires a choice of which parameters to use
Can pick based on sensible default values
- Random walk \(\theta=1\) often a good choice for autoregression with economic data
Usually apply a parameter selection method based on the data \[\theta=\widehat{\theta}(\mathcal{Y}_T)\]
Full forecasting method now a function of data again \[f(\mathcal{Y}_T)=f(\mathcal{Y}_T,\widehat{\theta}(\mathcal{Y}_T))\]
Many types of parameter selection methods exist
- Will discuss details in later classes
Selecting parameters not quite the same as “estimation” in statistics
- Care about \(f(\mathcal{Y}_T)\), not \(\theta\) itself

Implementation

Forecast using parametric methods, with default parameter selection rules
- Autoregressive models of order 1 and order 4
- Simple Exponential Smoothing

milkAR1selection<-Arima(condmilk,order=c(1,0,0),include.mean=FALSE) #Choose parameter of AR(1) model
    #Do not include added constant
milkAR1<-forecast(milkAR1selection,h=12) #Autoregression of order 1 forecast, 12 months out
milkAR4selection<-Arima(condmilk,order=c(4,0,0)) #Choose parameters of AR(4) model, including constant
milkAR4<-forecast(milkAR4selection,h=12) #Autoregression of order 4 forecast, 12 months out
milkexpsm<-ses(condmilk,h=12) #Simple Exponential Smoothing Forecast, 12 months out

#Plot the forecasts
autoplot(condmilk,series="Observed")+
  autolayer(milkAR1, PI=FALSE, series="AR(1) no constant")+
  autolayer(milkAR4, PI=FALSE,series="AR(4)")+
  autolayer(milkexpsm, PI=FALSE,series="Simple Exponential Smoothing")+
  ggtitle("Forecasts for Monthly Milk Stocks") +
  guides(colour=guide_legend(title="Series and Forecasts"))

Plot

Parameters Chosen

# alpha and l_0 from exponential smoothing
milkexpsm$model$fit$par

## [1]  0.99990 81.29198

# AR(1) Coefficient
milkAR1selection$coef

##       ar1 
## 0.9783295

# AR(4) Coefficients
milkAR4selection$coef

##         ar1         ar2         ar3         ar4   intercept 
##  0.98568158 -0.09175605 -0.11115186 -0.27093352 95.99392283

Implementation

Forecast using seasonal and drift methods
- Series appears highly seasonal, with minimal trend
- So expect better results from seasonal method

milksnaive<-snaive(condmilk,h=12) #Seasonal Naive forecast, 12 months out
milkdrift<-rwf(condmilk,drift=TRUE,h=12) #Drift forecast, 12 months out

#Plot the forecasts
autoplot(condmilk,series="Observed")+
  autolayer(milksnaive, PI=FALSE, series="Seasonal Naive")+
  autolayer(milkdrift, PI=FALSE,series="Drift")+
  ggtitle("Forecasts for Monthly Milk Stocks") +
  guides(colour=guide_legend(title="Series and Forecasts"))

Plot

Adding External Data

A forecast might be a function of series of interest AND other data, \(\mathcal{X}_T\)
- Related time series, Background information, etc
Method becomes \[\widehat{y}_{T+h}=f(\mathcal{Y}_T,\mathcal{X}_T,\text{other stuff})\]
Example: (Predictive) Regression
- Forecast is linear function of \(x_T\) \[\widehat{y}_{T+h}=f(\mathcal{Y}_T,\mathcal{X}_T,\theta)=\theta_0+\theta_1 x_T\]
More generally, may use subjective or judgmental assessment to come up with a number
Judgemental forecast: think real hard and come up with a number \[\widehat{y}_{T+h}=f(\mathcal{Y}_T,\text{whatever you're thinking})=???\]
Expert forecast: ask someone else to think real hard and come up with a number \[\widehat{y}_{T+h}=f(\mathcal{Y}_T,\text{whatever they're thinking})=???\]

Random methods

Some forecasting methods involve randomness
Let \(U\) be a random variable, with distribution \(F(u)\)
A randomized method takes the random variable as an input, so its output is also random
E.g. Random Guess: pick a number at random, without using the data \[f(\mathcal{Y}_T,U)=U\]
- A random guess is a random generalization of trivial method
- A randomized version of any other procedure can be made by adding \(U\) to the output
Not super common, but one part of some more complicated methods

Choices: Finding a “Good” method

Now we know we CAN use any of these methods to forecast
What SHOULD we use?
- Depends on application
Want a good guess because bad guesses will lead to bad decisions
- Sales forecasts too high or low both lead to incorrect inventory decisions, lower profit
- Financial market forecasts too high or low lead to bad buy/sell decisions, lower profit
- Missed economic indicator forecasts can lead to inappropriate policy, worse outcomes
Ideal forecast: exact true value \(y_{T+h}\)
- Problem: value isn’t known
Next best choice: something very close to \(y_{T+h}\)
- Problem: also not known
- Problem: what is close?
Need a measure of how bad decision will have turned out to be given an incorrect forecast
- A “Loss Function”

Loss Functions

Mathematically, a loss function \(\ell(.,.)\)
- Takes true value \(y\) and forecast value \(\widehat{y}\)
- Outputs a number measuring how bad the guess was
After truth is known, says how bad forecast was
Acts like (dis)utility of outcome
- If used to utility functions from economics, this is negative of utility
Usual property
- \(y=\underset{\widehat{y}}{\arg\min}\ell(y,.)\) for all \(y\)
- A perfect forecast is alway (weakly) better than other guesses
If above holds, can use normalized version
- \(\ell(y,y)=0\) for any \(y\). 0 Loss for exactly correct guesses
- \(\ell(y_1,y_2)\geq 0\) for any \(y_1,y_2\). Incorrect guesses are (weakly) worse
If profit \(\pi(y,\widehat{y})\) is maximized by correct forecasts
- \(\ell(y,\widehat{y})=\pi(y,y))-\pi(y,\widehat{y})\) is equivalent loss function

Examples

Common loss functions \(\ell(y,\widehat{y})\) use measures of absolute or relative distance
Absolute Error \[|y-\widehat{y}|\]
Squared Error \[(y-\widehat{y})^2\]
Absolute Percent Error: Requires non-zero values of \(y\) \[\left|100\frac{(y-\widehat{y})}{y}\right|\]
Absolute Scaled Error: Compares error to average error of a random walk forecast \[\left|\frac{(y-\widehat{y})}{\frac{1}{T-1}\sum_{t=2}^T|y_t-y_{t-1}|}\right|\]

Evaluating Loss

To see how forecast method performed in past, can evaluate loss within data \(\mathcal{Y}_T\)
Sample mean of loss function common way to describe past performance \[\frac{1}{T-h}\sum_{t=1}^{T-h}\ell(y_{t+h},f(\mathcal{Y}_{t}))\]
For Squared Error, this is called MSE, for Absolute Error, MAE, for Abolute Percent Error, MAPE, Absolute Scaled Error MASE
- Usually take square root of MSE to get RMSE: Units same as series
Answers question: “How much loss would I have incurred (on average) had I used this method in the past?”
Not same as “How much loss will I incur if I use this method in the future?”

Evaluating Parametric Methods

With parameterized methods, common to use different data for forecast rule and parameter selection rule \[\widehat{y}_{t+h}=f(\mathcal{Y}_t,\widehat{\theta}(\mathcal{Y}_T))\]
Predicted value for \(y_{t+h}\) that you would predict at \(t\) using the forecast rule chosen using all the data up to \(T\)
- Mostly done for computational reasons: methods to calculate \(\widehat{\theta}(\mathcal{Y}_T)\) can be very slow
Values \(y_{t+h}-f(\mathcal{Y}_t,\widehat{\theta}(\mathcal{Y}_T))\) are called residuals
- \(\frac{1}{T-h}\sum_{t=1}^{T-h}\ell(y_{t+h},f(\mathcal{Y}_t,\widehat{\theta}(\mathcal{Y}_T)))\) is residual error
- Also given name MSE, MAE, MAPE, etc
Residuals are a measure of error, but do not describe loss from a feasible procedure
To compare a feasible measure, also use only past data for forecasting rule
- Predict \(\widehat{y}_{t+h}=f(\mathcal{Y}_t,\widehat{\theta}(\mathcal{Y}_t))\)
- Errors and loss from this approach called “evaluation on rolling forecast origin”
- Calculate via tsCV in forecast package
- Can be slow: calculate \(\widehat{\theta}(\mathcal{Y}_t)\) many times

Evaluation

Construct rolling forecast origin AR4 models and evaluate 1-month errors

far4 <- function(x, h){forecast(Arima(x, order=c(4,0,0)), h=h)}
AR4forecasterrors<-tsCV(condmilk,far4,h=1)
#Rolling Forecast Error RMSE and MAE
RMSEfc<-sqrt(mean((AR4forecasterrors)^2,na.rm = TRUE))
MAEfc<-mean(abs(AR4forecasterrors),na.rm = TRUE)
#Compare RMSE and MAE on residuals
RMSEres<-sqrt(mean((milkAR4selection$residuals)^2,na.rm = TRUE))
MAEres<-mean(abs(milkAR4selection$residuals),na.rm = TRUE)

Method	RMSE	MAE
Residual	12.7383456	8.9054659
Rolling	15.4255888	10.2356219

Rolling window errors are larger than residual errors
Former reflect what you would have gotten from running method in the past

Train vs Test Set

Simple compromise: Split \(\mathcal{Y}_T\) into \[\mathcal{Y}_{Train},\mathcal{Y}_{Test}\] \[\mathcal{Y}_{Train}=\{y_t\}_{t=1}^{S},\mathcal{Y}_{Test}=\{y_t\}_{t=S+1}^{T}\]
- where \(S\) is split point
Compute error as \[\frac{1}{T-S}\sum_{t=S+1}^{T}\ell(y_{t},f(\mathcal{Y}_{\min\{t-h,S\}},\widehat{\theta}(\mathcal{Y}_{Train})))\]
Allows fast computation: compute \(\widehat{\theta}(\mathcal{Y}_{Train})\) only once
Performance measure describes feasible method, since test set not used

Train vs Test Set

Compute test set accuracy for milk forecasts by many measures

milktrain<-window(condmilk,end=c(1977,12)) #Until 1978
milktest<-window(condmilk,start=c(1978,1)) #After 1978                  
milkAR4train<-Arima(milktrain,order=c(4,0,0))
milkAR4forecast<-forecast(milkAR4train,h=36)
#Test set RMSE and MAE
RMSE<-sqrt(mean((milkAR4forecast$mean-milktest)^2))
MAE<-mean(abs(milkAR4forecast$mean-milktest))

Method	RMSE	MAE
Test Set Error	17.5138905	14.6526401

#Plot results
autoplot(milktrain, series="Training Set Data")+
  autolayer(milktest,series="Test Set Data")+
  autolayer(milkAR4forecast, PI=FALSE,series="Forecast")+
  ggtitle("Forecasts for Monthly Milk Stocks") +
  guides(colour=guide_legend(title="Series and Forecasts"))

Plot

The problem of induction

Now that we can measure forecasting performance, what do we know about how to forecast?
- Not much!
Problem:
“Past performance does not necessarily predict future results”
- U.S. Securities and Exchange Commission mandatory warning, paraphrasing Scottish Enlightenment philosopher David Hume

The problem of induction

No matter how good the empirical performance of the method, or how much data, cannot say anything at all about what will happen on new data without additional logical premises
Will sun rise tomorrow?
- Has done so every recorded day in the past
- Premise that things that have always happened before will continue to happen might be true
- But can’t be justified by the fact that it’s been true in the past
Circularity of reasoning best expressed by considering opposite belief (cf Aaronson (2013) p229)
There is “a planet full of people who believe in anti-induction: if the sun has risen every day in the past, then today, we should expect that it won’t. As a result, these people are all starving and living in poverty. Someone visits the planet and tells them,”Hey, why are you still using this anti-induction philosophy? You’re living in horrible poverty"
- They respond; “Well, it never worked before…”
Clearly, we are going to need some additional principles

Next class

Principles of forecasting
- No “right” answer
- Multiple reasonable approaches
Submit first problem sets

Methods and Motivation

73-423 Forecasting for Economics and Business

Building a Forecast

Simple Forecast Methods

Implementation

Plot

Methods with parameters

Examples of Parametric Methods

Structured parametric models

Choosing parameters

Implementation

Plot

Parameters Chosen

Seasonal and trending data

Implementation

Plot

Adding External Data

Random methods

Choices: Finding a “Good” method

Loss Functions

Examples

Evaluating Loss

Evaluating Parametric Methods

Evaluation

Train vs Test Set

Train vs Test Set

Plot

The problem of induction

The problem of induction

Next class