See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/311945797

Revenue Forecasting using Holt–Winters Exponential Smoothing

Article · December 2016

CITATIONS READS

20 6,861

4 authors:

Md. Habibur Rahman Umma Salma

Jahangirnagar University Jahangirnagar University
15 PUBLICATIONS   41 CITATIONS    2 PUBLICATIONS   25 CITATIONS   

SEE PROFILE SEE PROFILE

Md. Moyazzem Hossain Md Tareq Ferdous Khan

Jahangirnagar University Cincinnati Children's Hospital Medical Center
139 PUBLICATIONS   754 CITATIONS    20 PUBLICATIONS   50 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Finding a convenient kernel for stock market data. View project

Project Impossible: International Multi-Disciplinary Publishing Team View project

All content following this page was uploaded by Md. Moyazzem Hossain on 20 March 2017.

The user has requested enhancement of the downloaded file.

 Research & Reviews: Journal of Statistics
ISSN: 2278-2273(online), ISSN: 2348-7909(print)
Volume 5, Issue 3
www.stmjournals.com

Revenue Forecasting using Holt–Winters

Exponential Smoothing
Md. Habibur Rahman1,*, Umma Salma2, Md. Moyazzem Hossain3,
Md. Tareq Ferdous Khan4
1,3,4
Department of Statistics, Jahangirnagar University, Dhaka, Bangladesh
2
Research Evaluation Associates for Development Ltd., Dhaka, Bangladesh

Abstract
A lot of time series demonstrate seasonal behavior with trend, such type of series was monthly
revenue (in crore) of Bangabandhu Multipurpose Bridge. The seasonal forecasting with trend
issue was considerable importance. The research work focus on the analysis of seasonal time
series data using additive and multiplicative seasonal model of Holt–winters method and
forecast the monthly revenue (in crore) using best model—additive Holt–Winters exponential
smoothing up to January 2021.

Keywords: Forecast, levels, seasonality, trend, root mean squared, Bangladesh

*Author for Correspondence E-mail: habib.drj@gmail.com

INTRODUCTION approach to forecasting seasonal time series.

Forecasting was most often part of a larger Winters method and Fourier series analysis are
process of planning and managing, and a versatile methods because the methods model
forecast was necessary to provide accurate the level, trend, and seasonality of a time
estimates of the future for the larger process. series (DeLurgio SA, 1998) [5]. Puah YJ,
The all of forecasting methods were classified Huang YF, Chua KC, Lee TS. (2016) [6] were
under the types—time series univariate modeled the Rainfall series using additive
methods, causal or multivariate methods, Holt–Winters method to examine the rainfall
qualitative or technological, and other pattern in Langat River Basin, Malaysia. The
quantitative methods. Holt–Winters exponential smoothing is one of
the most popular methods to forecast the
The Holt–Winters method is a statistical different time series. The main objective of the
forecasting method in time series univariate research work was fit the additive Holt–
methods. The forecasting engages make Winters exponential smoothing model, check
projection about future performance on the the accuracy measures, and forecast the
basis of past and recent data. Dasgupta SS, monthly revenue (in crore) of the
Mahanta P, Roy R, Subramanian G. (2014) [1] Bangabandhu multipurpose bridge.
forecast industry big data with Holt–Winters
method. Exponential smoothing models and DATA DESCRIPTION
auto regressive moving average model The monthly revenue (in crore) from July
compared to comprehend which method is 1998 to July 2016 was plotted and shown in
more adapted to model the temperature Figure 1. The mean revenue found to be
behavior in Caserta, Italy (Guizzi G, Silvestri 16.917 with standard deviation 8.9375. The
C, Romano E, Revetria R, 2015) [2]. quantile—quantile plot and box plot was
presented in the Figure 2. The monthly
Taylor JW (2003) [3] applied the Holt– revenue data normally distributed. From the
Winters method to forecast short-term above time series plot, it was evident that the
electricity demand. Holt (2004) [4] extended seasonality and trend was present in the series.
the exponential weighed moving averages to The seasonal length of the series was twelve
allow trend and seasonal variation. Holt– (Table 1).
Winters exponential smoothing is a popular

RRJoST (2016) 19-25 © STM Journals 2016. All Rights Reserved Page 19
Revenue Forecasting using Holt–Winters Exponential Smoothing Rahman et al.

METHODOLOGY
The monthly revenue (in crore) of
Bangabandhu Multipurpose Bridge collected
from Bangladesh Bridge Authority, Bridges
Division, Ministry of Road Transport and
Bridges for July 1998 to July 2016. The toll
was collected at Bangabandhu Multipurpose
Bridge which stands on the river Jamuna
between the district Sirajgonj and Tangail.

Exponential smoothing methods give larger

weights to more recent observations, and the
weights decrease exponentially as the
observations become more distant. These
methods are most effective when the
parameters describing the time series are
changing slowly over time. Holt–Winters
Methods are applicable when both trend and
seasonal pattern present. Two Holt–Winters
methods were designed for time series that
exhibit linear trend. Additive Holt–Winters
method used for time series with constant
(additive) seasonal variations and
Multiplicative Holt–Winters method used for
time series with increasing (multiplicative)
seasonal variations.

Table 1: Summary statistics.

Monthly Revenue (in crore)
Mean (in crore) 16.9170
Standard deviation (in crore) 8.9375
Fig. 2 (a) Q-Q Plot and (b) Box Plot for
Kurtosis 0.9395
Monthly Revenue (in crore).
Skewness 0.4038
Coefficient of variation 0.5283
Holt–Winters Method
Minimum (in crore) 3.7900
The Holt–Winters method was originally
Maximum (in crore) 37.7200
developed by Winters and involves estimating
Count 217.0000
three smoothing parameters associated with
the level, trend and seasonal factors
Time Series Plot (Bermudez JD, Segura JV, Vercher E, 2006)
[7]. Holt–Winters method is an exponential
35

smoothing approach for handling seasonal

Revenue(in Crore)

data.
25

Additive Holt–Winters Method

The additive Holt–Winters method is
15

presented in the following equations.

yT  0  1t  snT  T
5

Estimate of the level at time T

lT    yT  snT  L   1    lT 1  bT 1 
2000 2005 2010 2015
Estimate of the growth rate (or trend) at time T
Time(in Month) bT    lT  lT 1   1    bT 1 , 0   ,   1
Fig. 1: Time Series Plot of Monthly Revenue Estimate of the seasonal factor at time T
(in crore).

RRJoST (2016) 19-25 © STM Journals 2016. All Rights Reserved Page 20
Research & Reviews: Journal of Statistics
Volume 5, Issue 3
ISSN: 2278-2273(online), ISSN: 2348-7909(print)

snT    yT  lT   1    snT  L where, 0    1 e 2

MSE 
t
Mean Squared Error :
p- Step ahead forecast made at time T n
yˆT  p T   lT  pbT  snT  p  L where, p  1, 2,... e 2

RMS 
t
Root Mean Squared :
Where, n
Mean Absolute Scaled Error:
Initial level=l0   0  intercept n

Initial growth rate=b0  1  Slope 1 n yt  yt

ˆ

1 n
MASE  
n t 1 1 n
t 1

ST  yT  yˆT , and Si   S2k 1

L k  i 1
 yt  yt 1
n  1 t 2
L  No.of seasons in a year ESS RSS
R2   1
Initial seasonal factors  sni  L  Si  , i  1, 2, ..., L TSS TSS
 y  yˆt 
Percentage Error : PEt  t  100
yt
Multiplicative Holt–Winters Method
The multiplicative Holt–Winters method is  n  1 RSS  n  1 2
Adjusted R 2 : R 2  1   R
presented in the following equations.  n  k  TSS  n  k 
yT   0  1t   snT   T  PE
Mean Percentage Error : MPE 
t

Estimate of the level at time T n

lT    yT snT  L   1    lT 1  bT 1   PE
Mean Absolute Percentage Error : MAPE 
t

Estimate of the growth rate (or trend) at time T n

bT    lT  lT 1   1    bT 1 ,0   ,   1 The TSS, ESS, and RSS mean total sum of
square, explained sum of square and residual
Estimate of the seasonal factor at time T
sum of square respectively.
snT    yT lT   1    snT  L where, 0    1
p- Step ahead forecast made at time T RESULT AND DISCUSSIONS
yˆT  p T    lT  pbT   snT  p  L , p  1, 2,... The trend and the seasonal pattern were
Where, present in the time series (monthly revenue in
crore). The Holt–winters method is applicable
Initial level=l0   0  intercept to forecast the future state. The Holt–Winters
was fitted to forecast monthly revenue (in
Initial growth rate=b0  1  Slope
crore) for the Bangabandhu Multipurpose
1 n
ST  yT yˆT and Si    S2k 1
L k  i 1
Bridge of Bangladesh. Finally, the equations
take the forms and the fitted models with
L  No.of seasons in a year estimated parameters are shown in following.
L
Normalized Constant = CF  L  S 
i 1
i Additive Holt–Winters Model
yT   0  1t  snT   T
Initial seasonal factors  sni  L  Si   CF 
l0   0  35.2931090 and b0  1  0.1192485
where, i  1, 2, ....., L
ST  yT  yˆT where, t  1, 2,..., n and n  217
1 n
Model Performance
Goodness of fit is the measure of the accuracy
Si    S2k 1
L k i 1
of the forecasted model to actual value. The where, L  No.of seasonsin a year=12
model forecast accuracy was measured by the
following criteria. S1  0.3680168 S7  1.2345571
S2  0.3639068 S8  1.0573878
MAD 
e t
Mean Absolute Deviation :
n S3  1.7726623 S9  0.1873891
Sum Squared Error : SSE   et2 S4  1.652559 S10  2.0314614

RRJoST (2016) 19-25 © STM Journals 2016. All Rights Reserved Page 21
Revenue Forecasting using Holt–Winters Exponential Smoothing Rahman et al.

S5  0.1115694 S11  3.0176877 S11  1.1666387 S12  1.1467173

S6  0.6515678 S12  1.9456332

sni  L  Si  where, i  1, 2, ....., L
sni  L  Si where, i  1, 2, ....., L
lT    yT snT  L   1    lT 1  bT 1 
lT    yT  snT  L   1    lT 1  bT 1 
where,   0.1766912
where,   0.5370757 bT    lT  lT 1   1    bT 1
bT    lT  lT 1   1    bT 1 where,   0.01577167
where,   0.005237878 snT    yT lT   1    snT  L
snT    yT  lT   1    snT  L where,   0.3461822
where,   0.4206121
p  step ahead forecast made at time T
yˆT  p T    lT  pbT   snT  p  L where, p  1, 2,...
p  step ahead forecast made at time T
n
yˆT  p T   lT  pbT  snT  p  L where, p  1, 2, ... SSE    yT  yˆT T  1   762.77962
t 1
n
SSE    yT  yˆT T  1   691.3605 The accuracy measures—Sum Square Error,
t 1 Mean Sum Square Error, Root Mean Squared,
Mean Percentage Error, Mean Absolute
Multiplicative Holt–Winters Model Percentage Error, Mean Absolute Deviation,
yT    0  1t   snT   T Mean Absolute Scaled Error of the additive
2

l0  0  32.5523721 and b0  1  0.1640977 Holt–Winters method are lower than R and

Adjusted R2 higher than the multiplicative
ST  yT yˆT where, t  1, 2,..., n and n  217
Holt–Winters method (Table 2).
1 n
Si   S2k 1
L k i 1 The fitted seasonal factor, growth rate (or
where, L  No.of seasonsin a year=12 trend), level and forecasted revenue were
plotted against time (month) presented in
S1  1.0962413 S2  1.0634873
Figure 3. The different types of forecast
S3  1.1492499 S4  1.0092294 accuracy measures of additive Holt–Winters
methods such as sum square error, root mean
S5  1.0478964 S6  1.0121130 squared, mean absolute percentage error, mean
2
S7  0.9686139 S8  1.0575681 absolute deviation, and R are 691.3605056,
1.836434118, 5.899982004, 1.075149359,
S9  1.0117472 S10  1.1068369 0.972069439 respectively (Table 3).

Table 2: Forecast Accuracy Measures.

Accuracy Measures Holt–Winters Additive Model Holt–Winters Multiplicative Model

SSE 691.3605056 762.7796266

MSE 3.372490271 3.720876227

RMS 1.836434118 1.928957290

R2 0.972069439 0.968063769
2
Adjusted R 0.981693889 0.955882545

MPE 0.542534439 0.967230691

MAPE 5.899982004 6.771234123

MAD 1.075149359 1.148890118

MASE 0.520340000 0.556028300

RRJoST (2016) 19-25 © STM Journals 2016. All Rights Reserved Page 22
Research & Reviews: Journal of Statistics
Volume 5, Issue 3
ISSN: 2278-2273(online), ISSN: 2348-7909(print)

30
Fitted Value Holt-Winters Filtering
xhat
20
5 10
30
level
20
5 10
trend
0.08
0.04
2
season
1
0
-1
-2

2000 2005 2010 2015

Time(in Month)

Fig. 3: Fitted Seasonal Factor, Growth Rate (or trend), Level and Forecasted Revenue Plot against
Time (Month) for Additive Holt–Winters Model.

Fig. 4: Actual revenue (Black colour) and Forecasted Revenue (in crore) with Confidence Limit Plot
against Time (month) using Holt–Winters Additive Model.

RRJoST (2016) 19-25 © STM Journals 2016. All Rights Reserved Page 23
Revenue Forecasting using Holt–Winters Exponential Smoothing Rahman et al.

Table 3: Monthly Forecasted Revenue (in SUMMARY AND CONCLUSION

crore) using Additive Holt–Winters Method. The forecasting accuracy measures of the
Year Month Forecasted Revenue (in Crore) multiplicative Holt–Winters method were
2016 Aug 35.78037 higher than the additive Holt–Winters model.
2016 Sep 35.16770 On the basis of the accuracy measures the
2016 Oct 37.42352
monthly revenue (in crore) was forecasted by
2016 Nov 34.11754
2016 Dec 35.77778
the additive Holt–Winters method.
2017 Jan 35.35703
2017 Feb 34.89329 The forecasted revenue (in crore) were
2017 Mar 37.30448 presented where the forecasted value in
2017 Apr 36.17896 December 2016, June 2017, January 2018,
2017 May 38.51706 December 2019, and October 2020 were
2017 Jun 39.62253 35.77778, 39.62253, 36.78801, 40.07073,
2017 Jul 38.66972 43.14745 respectively. The forecasted revenue
2017 Aug 37.21136 was also upward trend.
2017 Sep 36.59868
2017 Oct 38.85450
The coefficient of variation of the monthly
2017 Nov 35.54853
revenue (in crore) and monthly forecasted
2017 Dec 37.20876
revenue (in crore) using additive Holt–Winters
2018 Jan 36.78801
method was 0.5283 and 0.06048151
2018 Feb 36.32427
respectively. The forecasting method—
2018 Mar 38.73547
2018 Apr 37.60994
additive Holt–Winters method be the superior
2018 May 39.94804
than the multiplicative Holt–Winters method
2018 Jun 41.05351
(Figure 4).
2018 Jul 40.10071
2018 Aug 38.64234 REFERENCES
2018 Sep 38.02966 1. Dasgupta SS, Mahanta P, Roy R,
2018 Oct 40.28548 Subramanian G. Forecasting Industry Big
2018 Nov 36.97951 Data with Holt Winter’s Method from a
2019 Dec 38.63975 Perspective of In-Memory Paradigm. On
2019 Jan 38.21900 the Move to Meaningful Internet Systems:
2019 Feb 37.75526 OTM 2014 Workshops, Lecture Notes in
2019 Mar 40.16645 Computer Science, 2014;8842:80-85.
2019 Apr 39.04092 DOI: 10.1007/978-3-662-45550-0_11.
2019 May 41.37902 2. Guizzi G, Silvestri C, Romano E, Revetria
2019 Jun 42.48449 R. A comparison of forecast models to
2019 Jul 41.53169 predict weather parameters. Advances in
2019 Aug 40.07332 Energy and Environmental Science and
2019 Sep 39.46065 Engineering. 2015;88-96.
2019 Oct 41.71646 3. Taylor JW. Short-term electricity demand
2019 Nov 38.41049 forecasting using double seasonal
2019 Dec 40.07073 exponential smoothing. Journal of the
2020 Jan 39.64998 Operational Research Society.
2020 Feb 39.18624
2003;54(8):799–805.
2020 Mar 41.59743
DOI:10.1057/palgrave.jors.2601589.
2020 Apr 40.47190
4. Holt CC. Forecasting seasonals and trends
2020 May 42.81000
by exponentially weighted moving
2020 Jun 43.91548
averages. Int. J. Forecasting.
2020 Jul 42.96267
2004;20(1):5–10.
2020 Aug 41.50430
2020 Sep 40.89163
5. DeLurgio SA. Forecasting principles and
2020 Oct 43.14745
applications, New York: McGraw Hill;
2020 Nov 39.84147
1998. pp 205.
2020 Dec 41.50171
2021 Jan 41.08096

RRJoST (2016) 19-25 © STM Journals 2016. All Rights Reserved Page 24
 Research & Reviews: Journal of Statistics
Volume 5, Issue 3
ISSN: 2278-2273(online), ISSN: 2348-7909(print)

6. Puah YJ, Huang YF, Chua KC, Lee TS.

River catchment rainfall series analysis Cite this Article
using additive Holt–Winters method. Md. Habibur Rahman, Umma Salma,
Journal of Earth System Science. Md. Moyazzem Hossain, Md. Tareq
2016;125(2):269–283. Ferdous Khan. Revenue Forecasting
DOI: 10.1007/s12040-016-0661-6. using Holt–Winters Exponential
7. Bermudez JD, Segura JV, Vercher E. Smoothing. Research & Reviews:
Improving demand forecasting accuracy Journal of Statistics. 2016; 5(3):
using nonlinear programming software. 19–25p.
Journal of the Operational Research
Society. 2006;57:94–100.

RRJoST (2016) 19-25 © STM Journals 2016. All Rights Reserved Page 25

View publication stats