View
2
Download
0
Category
Preview:
Citation preview
Product sales forecasting using online reviews and historicalsales data: A method combining the bass model and
sentiment analysis
Zhi-Ping Fan, Yu-Jie Che, Zhen-Yu Chen
Department of Information Management and Decision Sciences, School of Business Administration, Northeastern University, Shenyang 110167, China
Journal of Business Research 74 (2017) 90–100 [SSCI]
Speaker Yejin Kim
Date 5th September 2017
1
NEMO English Seminar
• Introduction
• Theoretical Background
• Mathematical Background
• Methology
• Data and results
• Conclusion
2
Contents
• Key idea
- Online reviews have a significant influence on product sales.
• Background.
- advertisements and the mass media, among other factors influence consumers' purchasing decisions.
- Word of mouth (WOM) is considered one of the most important factors influencing the purchasing decisions of consumers.
• Propose of paper
- We propose a bass Model considering the sentiments expressed in the content of online reviews For Sales forecasting
3
Introduction
• Product sales forecasting based on online review data
- Asur and Humberman (2010) : the sentiments extracted from the chatter of Twitter.com about movie improve box office sales forecasting
• Product sales forecasting using the Bass model
-Dellarocas et al. (2007) : developed a Bass model based on the revenue forecasting model using online ratings and the number of posted reviews
• Difference
- we extract the sentiment index from the content of online reviews, rather than ratings, and use it to extend the Bass and Norton model.
4
Theoretical Background
• Bass Model
(1)
(2)
(3)
• Bass Model differential equation
5
Mathematical Background
m -- the potential market (the ultimate number of adopters)p -- coefficient of innovation, External influenceq -- coefficient of imitation, internal influence f(t) -- the portion of M that adopts at time t.F(t) -- the portion of M that have adopted by time t
S(t) -- cumulative adopters (or adoptions) at t
• Bass & North Model
- Considered with next generation.
- Each generation has its potential market
• ith generation’s cumulative fraction
of adopters in time period t
6
Mathematical Background
𝑚𝑖 -- potential for the ith generation𝜏𝑖 -- the time at which ith generation is introduced𝑆𝐼 𝑡 -- Sales of the ith generation in time period t
• Step 1. Data collection and Preprocessing
-Collection Period : 2007.07 – 2014.12
-Collection Review site : Car Website ‘Bitauto’ of china
7
Methodlogy
Product Collection Period Time Period(per 3Month) Total Review
Elantra 2007.07-2014.12 30 1407
Elantra-1 2008.04-2014.12 27 2524
Elantra-y 2012.08-2014.12 10 368
• Step 2. Sentiment Extraction & Model Building
> Sentiment Extraction - NB(naïve bayes) algorithm
• Used Sentiment dictionary : CNKI sentiment dictionary
• Sentiment Categories : Ci with i∈{+,−} → negative, positive
8
Methodlogy
Wt -- sentiment index in time period t Wtk -- the percentage of words corresponding to category i Of the emotion words in the kth reviewDk -- -- set of emotional words in kth review i – positive or negative categoriest – time periodh – the number of reviews in the time periodc – positive or negative categories value
• Step 2. Sentiment Extraction & Model Building
> Focecasting model – Extended Bass model
- the cumulative sales by the end of time period t
- a function of the online review
sentiment index in the bass-emotion model
9
Methodlogy
=𝑞0𝑒𝛾𝑤𝑡
1 + ൘𝑞0(𝑒𝛾𝑤𝑡 − 1)𝑞𝑚
𝑞0 -- minimum of q𝑞𝑚 -- maximum of q𝛾 -- constant that controls the steepness of the S-curve
= proportional increase of the q in one unit of time (Verhulst, logistics equation)
𝑃 𝑡 =𝑝(0)𝑒𝑟𝑡
1 + 𝑝(0)(𝑒𝑟𝑡 − 1)/𝐾
• Step 2. Sentiment Extraction & Model Building
> Focecasting model – Extended Norton model
- the cumulative sales by the end of time period t
- a function of the online review
sentiment index in the Norton-emotion model
10
Methodlogy
𝑞𝑖0 -- minimum of qi
𝑞𝑖𝑚 -- maximum of qi
𝛾𝑖 -- constant that controls the steepness of the S-curve ith generation. = proportional increase of the qi in one unit of time
- Parameter Result
- Forecasting results
- Forecasting data
11
Data and results - Bass-emotion Model
R-square RMSE
0.9987 1.4910
- Parameter Result
- Statistics of Norton-emotion model
12
Data and results – Norton-emotion Model
- Forecasting results
13
Data and results – Norton-emotion Model
- Forecasting data
14
Data and results – comparison of MAPE
15
Conclusion
• a forecasting model that combines the Bass/Norton model and sentiment analysis techniques is proposed.
• this paper extends the Bass model by analyzing sentiments expressed in online reviews
• the proposed models exhibit lower forecasting errors than the comparative models
• Since m is a constant, The range of the predicted value is limited.
• Step 1. Data collection
16
Appendix – Step 1. Data collection
• Step 1. Data collection
17
Appendix – Step 1. Data collection
Collectable Sale Period
Collectable Review Period
Both Collectable Period
Elantra 2006.04-2014.12
2007.07-2015.02
2007.07-2014.12
Elantra-1 2008.04-2014.12
2008.04-2015.02
2008.04-2014.12
Elantra-y 2012.08-2014.12
2012.04-2015.03
2012.08-2014.12
Recommended