7 - demand forecasting

Question 1

why is demand forecasting used?

Answer 1

forecasting is an initial step in planning and decision making in various domains such as marketing and sales, logistics, financial investment, policy formulation, etc

having more accurate image of future would lead to better decision makings, and consequently higher profits/better service etc

all models are wrong, but some are useful - george box

even best statistical or machine learning forecasting model cannot perfectly represent the real world

Question 2

what are the main characteristics of forecasts?

Answer 2

forecasts always inaccurate and should thus include both expected value of forecast and measure of forecast error. small error preferred.

long term forecasts are usually less accurate than short-term forecasts. forecast error increases for longer forecasts

aggregate forecasts usually more accurate than disaggregate.

Question 3

what are the 4 types of forecasting methods?

Answer 3

1. qualitative - primarly subjective, rely on judgement. used most when little data.
2. time series - only uses historical demand. best w stable demand
3. causal - relationship between demand and some other factors
4. simulation - intimate consumer choices that give rise to demand. point is to try what if analysis.

Question 4

what is the objective of forecasting?

Answer 4

filter out random component and estimate systematic component.

Question 5

how is observed demand defined?

Answer 5

observed demand = systematic component + random component

random component = forecast error

Question 6

what is a systematic component’s level?

Answer 6

baseline value or average around which data fluctuates

represents central tendency of time series data

doesn’t have to stay the same, can be updated.

Question 7

what is a systematic component’s trend?

Answer 7

long run behaviour of data

increasing, decreasing, constant

need a few periods of data to identify a trend

Question 8

what does it mean for a systematic component to have seasonality?

Answer 8

periodic fluctuations that repeat themselves at fixed time intervals

occur due to periodic nature of data

period can be hourly, weekly, quarterly etc

Question 9

what is a systematic component’s cycle?

Answer 9

long term fluctuations

occur in long run periods - usually over several years

there is a reason for it

represents ups and downs of economy or of a specific industry

to properly isolate cycle from other components, you need data over decades

for practical purposes, they are usually ignored. in that case the cycle (if existent) will be considered as part of other components.

not relevant to operational problems, more for macro problems.

doesn’t have fixed term like others, more about the start/end points

Question 10

what are some examples of time series approaches?

Answer 10

naive approach - observe demand for period and imagine it doesn’t change for next.

moving average

exponential smoothing

ARIMA models

Question 11

how does a moving average work?

Answer 11

simple moving average — level only (no obvious trend or seasonality)

other extensions like exponential moving average and weighted moving average

Question 12

how can exponential smoothing be used for time series approaches?

Answer 12

simple exponential smoothing — level only

holt’s model (double exponential smoothing with additive trend) — level and trend but no seasonality

winter’s model (triple exponential smoothing) — level and trend and seasonality

Question 13

when is the moving average used?

Answer 13

used when demand has no observable trend or seasonality. in this case, systematic component of demand = level.

can catch some trend but not best when there is obvious trend in the data.

level should be updated as more info becomes available

level in t is average demand over last n periods

forecast for next period = estimate of level for period t

Question 14

how is the moving average updated?

Answer 14

after observing demand for period t+1, revise estimates to include it.

forecast for all future periods at period t is the same, and based on estimate of level at period t.

moving average corresponds to giving last n period of data equal weight when forecasting and ignoring all data older than this new moving average.

moving average becomes less responsive as N increases.

Question 15

when is simple exponential smoothing appropriate?

Answer 15

when demand has no observable trend or seasonality. similar to moving average, systemic component of demand = level.

initial estimate or level L0, assumed to be first observed value.

L0 = D1. after observing Dt+1 for period t+1, revise estimate to
Lt+1 = alphaDt+1 + (1-alpha)Lt).

alpha = smoothing parameter, between 0 and 1.

Question 16

how is simple exponential smoothing updated?

Answer 16

revised value of level is a weighted avg of the observed value of the level Dt+1 in period t+1 and old estimate of Lt in period t.

forecast for all future periods at period t is same and based on estimate of level at period t

forecast that is more responsive to recent observations = higher alpha.

Question 17

what are the similarities between the moving average and simple exponential smoothing?

Answer 17

both appropriate for stationary series - no trend or seasonality. can still use them but weight must be on recent data points.

both depend on single parameter - N or alpha

if there is a trend in the data, both methods come with forecasts that lag behind observations

Question 18

what are the differences between moving average and simple exponential smoothing?

Answer 18

MA treats each data point in N past periods with same weight. ES puts more weight on recent data points

ES uses all past history to make forecasts, MA uses only past N periods

MA reqiures all N past data points as inputs. ES only requires last forecast and last observation as inputs. lsat forecast info has the ‘memory’ of all past data points

Question 19

what is trend-corrected exponential smoothing?

Answer 19

AKA Holt’s model

appropriate when demand is assumed to have a level and a trend in the systematic component, but no seasonality.

systematic component of demand = level + trend

for intitial estimate of level and trend, there are different approaches. L0 = D1, T0 = 0.

Question 20

what is the revised estimate in Holt’s model?

Answer 20

weighted average of observed value and old estimate

Question 21

why should forecast errors be considered?

Answer 21

forecast errors contain valuable info and must be analysed carefully for two reasons -

1. to determine whether current forecasting model is predicting the systematic component of demand accurately
2. to account for forecast error in contingency planse.g. determine certain amount of contingency capacity with local supplier to be used if demand exceeds quantity in far east supplier provides.

also important to see which forecasting approach provides lowest errors

Question 22

what should forecast error be, technically?

Answer 22

should technically be 0. if all -ve or +ve, you’re not capturing systematic component. need to reevaluate method.

size of error important for deciding contingency plans - do you have huge errors and you might suddenly have huge difference in forecast vs actual?

Question 23

what ways can forecast error be measured?

Answer 23

mean squared error

mean absolute deviation

Question 24

when should mean squared error be used to measure forecast error?

Answer 24

MSE penalises large errors much more significantly than small errors because they’re squared

good idea to use MSE to compare forecasting methods if cost of large error is much bigger than gains from very accurate forecasts.

most used and most popular method.

Question 25

when should the mean absolute deviation be used to measure forecast error?

Answer 25

appropriate choice when selecting forecasting methods if cost of error is proportional to size of error.

if -ve errors, gives you size measurement rather than direction.