Scribd
Upload
20 views6 pages

Market Forecasting m3 Notes

Uploaded by

Abi Fleischmann
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
20 views6 pages

Market Forecasting m3 Notes

Uploaded by

Abi Fleischmann
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 6

SMOOTHING METHODS:

Smoothing Methods Usefulness


- data visualization
- computing seasonal indexes and removing seasonality
- forecasting

Components of a Time Series


1. Level (always present) – defined as the average observation at a given time (average of
all values before)
2. Trend (steady increase/decrease over time) – line joining two points of data in set
3. Seasonality (pattern that repeats itself every season) – cyclical pattern observed (seasonal
number for a year data where data repeats monthly is 12)
4. Random noise (always present)

Modeling Principles
Time series Analysis
- reasonableness and parsimony
- goodness of fit (residual analysis)
Time series Forecasting
- forecast accuracy
- parsimony and reasonableness

Challenges with most Data Characteristics


- components change over time (change in trend/seasonality)
- predetermined structure of forecasting model is not useful after some time

SOLUTION  smoothing technique


- smooth the noise and uncover the pattern
- simple and adaptable (widely used in businesses)
- data driven (don’t worry about changes over time)
MOVING AVERAGES:

IDEA  forecast future points by using an average of several past points

Moving Average Method Usefulness


- time series visualization
- computing seasonal indexes
- forecasting

ADVANTAGES  simple, popular (academia and business)


KEY CONCEPT  width of window (indication of future values)

Types of Windows
Centered moving average
- based on a window centered around t
- more useful for data visualization
- t
Trailing moving average
- based on a window from time t and backwards
- t

Computing a Trailing MA
1. Choose window width (W)
2. For MA at time t, place window on time points (t-W+1,…,t)
3. Compute average of values in the window

Computing a Centered MA
- Odd width: center window on time t and average the values in the window
- Even width: take the two “almost centered” windows and average the values in them

* time-plot of the moving averages can help reveal he level and trend of a series, by filtering out
the seasonal and random components

Choosing Window Width (W)


- balance over- and under-smoothing
- if there is no seasonality, use narrow window (under-smoothing)
- cannot go under 2 ever
- for seasonal series, use W equal to number of seasons

MA not suitable if
1. series with trend
2. series with seasonality

Excel  chart tools > design > add chart element > trendline
EXAMPLE - 8:32 – 18:12 (XLMiner)
FORECASTING WITH EXPONENTIAL SMOOTHING:

Types of Exponential Smoothing


1. Simple exponential smoothing (for series with no trend or seasonality)
2. Holt’s method (with trend, no seasonality)
3. Winter’s method (with trend and seasonality)

Simple Exponential Smoothing


- assumption: series has only level () and noise
- noise is unpredictable
- level will stay put
- forecasts: estimated level at most recent time point
F t+ k =Lt
- adaptive algorithm: adjusts most recent forecast (or level) based on the actual data)
F t=α Y t +(1−α)F t−1
α = the smoothing constant (0< α ≤1 ¿
- initialization:
F 1=L1=Y 1

Why is it called Exponential Smoothing?


- weights decrease exponentially into the past
- compare to moving average

The Smoothing Constant α


- α determines how much weight is given to the past
- α  0: past observations have large influence on forecasts (over-smoothing)
- α =1: past observations have no influence over forecasts (under-smoothing)
- typical values: 0.1, 0.2
- trial and error – effect on visualization
- minimize RSME or MAPE of training data
2
- trailing MA with window W = simple exponential smoothing with α =
W +1

Benefits
- only need to store the last forecast and most recent observation
- unlike MA, it gives more weight to more recent observations
- simple to understand

XLMiner  time series > smoothing > exponential smoothing > default α =2
REMOVING TREND AND/OR SEASONALITY:

Approach 1: Regression
- to remove trend and/or seasonality, fit a regression model with trend and/or seasonality
- series of forecast errors should be de-trended and de-seasonalized

Approach 2: Differencing
- differencing means taking the difference between two observations
- differencing is a simple and popular operation for removing a trend and/or seasonality
from a time series
- lag-1 difference: y t − y t −1 for removing trend
- lag-M difference: y t − y M −1 for removing seasonality with M seasons
- double-differencing: difference the differenced series
- compared to residuals from regression with trend, differencing is useful for removing a
local/changing trend shape
- simple and popular operation for removing a trend and/or seasonality from a time series

EXAMPLE: 2:38 – (lag, de-trend, and de-season excel)

Approach 3: Ratio-to-Moving Average


- uses moving average to remove seasonality
- also generates seasonal indexes as a byproduct

Seasonal Indexes
- for series with M seasons:
S j = seasonal index for the jth season
* indicates the exceedance of Y on season j above/below the average of Y in a
complete cycle of seasons
- average of the M seasonal indexes is 1 (they must sum to M)

Computing Seasonal Indexes


1. construct the series of centered moving averages of span M
2. for each t, compute the raw seasonals = Y t /MA t
3. S j = average of raw seasonals belonging to season j (normalize to ensure that seasonal
indexes have average = 1)

Using Seasonal Indexes to De- and Re-Seasonalize


- de- seasonalized (= seasonally-adjusted) series:
DSY t=Y t/appropriate seasonal index
- if done appropriately, de- seasonalized series will not exhibit seasonality
- if so, examine for trend and fit a model
- this model will yield de-seasonalized forecasts
- convert forecasts by re-seasonalizing (multiply them by appropriate seasonal index)
ADVANCED EXPONENTIAL SMOOTHING:

IDEA  augment simple exponential smoothing by capturing a trend component

Holt’s Method (double exponential in XLMiner)


- forecasts = most recent estimate level + trend
F t+ k =Lt + k T t
- also known as linear trend model

Equation 1: Estimating the Level


Lt =α Y t +(1−α)(Lt −1 +T t−1 )
- when α =1, Lt =Y t
- when α =0 , naïve forecasting approach
- adjust the previous level by adding trend
- trend is not global, but changes over time

Equation 2: Estimating the Trend


T t=β ( Lt −Lt−1 ) + ( 1−β ) T t −1
- update previous trend using the difference between the most recent level values
- with global patterns, you want smaller β
- local trend: trend is allowed to vary adaptively over time (compare to regression-based
models) β=1

Choosing α and β
- default values
- minimize RMSE or MAPE (or other goodness-of-fit criterion) for training set
* danger of over-fitting – make sure chosen values are reasonable

What does β=0 mean?


1. The model is using the initial value of the trend
2. No updating was necessary in order to arrive at the. lowest possible RMSE
3. Technical problem with fitting Holt’s method to these data
4. There is no trend
WINTER’S METHOD:

Overview Equation
F t+ k =(L ¿ ¿t +k T t )S t+ k−M ¿
St = seasonal index of period t
M = number of seasons
- Forecasts take into account level, trend, and seasonality:
- multiplicative seasonality, additive seasonality also available
- If forecast horizon > one cycle of seasons, use the last set of estimated seasonal indexes

Level Equation
Yt
Lt =α +(1−α )(Lt−1 +T t−1)
S t− M

Trend Equation
T t=β ( Lt −Lt−1 ) + ( 1−β ) T t −1
- same as Holt’s

Seasonality Equation
Yt
St =γ + ( 1−γ ) St −M
Lt
- multiplicative

Technical Notes
- all three smoothing constants are between 0 and 1
- to obtain forecasts (in XLMiner) you must re-run the exponential smoothing on the non-
partitioned data otherwise you lose the information in the most recent observations
- initialization (technical):
* L1=Y 1 or L1=a from estimated model Y t =a+ bt
* T 1=Y 2−Y 1 or ( Y T −Y 1 ) /T (average overall trend)
* initial seasonal indexes = MA indexes (that were seen earlier)
- look for method/model that gives lowest MAP and choose that one

You might also like