forecasting swimsuit sales for the next month to assist ......forecasting swimsuit sales for the...
TRANSCRIPT
Forecasting swimsuit sales for the next month to assist inventory management for Heatwave
Team 4 :Jheng Kai-Ru (107078503) Adam Yu (107078506) Silvia Yang (107078507) Zoly Chang (107078509)
Heatwave is a B2C swimsuits seller on a Chinese e-commerce platform called TMall. They design and manufacture their own swimsuits
Sell on
Problem
Heatwave has limited knowledge toward decision on how many product to produce; also to put promotions on certain products. They mostly do it base on their experiences. Sometimes it works, sometimes it doesn’t.
Client
Goal
Using the historical sales data to forecast the sales of the next month to assist with their inventory management and production strategy.
Business Problem
Data Description: Data Constraints
● Data from Shen-Yi-Can-Mou is too short!● Order doesn’t match in order list and item list● Some products may have different product name, and
may have different product id● Two sources of data have different product id● We miss 1314 data when integrated them
Problems
Although the data are not very accurate, we think it will still be helpful for our forecasting : )
From 2017-05 to 2018-11Summer has higher sales
Data Description: After Preprocessing
Monthly Data
From 2017-04-22 to 2018-12-18Holiday has higher sales
Daily Data
Data Description: Original Data
● Data : Monthly and daily Sales data● Time Period :
Daily 2018/07/25 - 2018/12/18 Monthly 2017/09 - 2018/11
● Data Quality : good
● Data : Daily data, contain order list and item list
● Time Period : 2017/04/22 - 2018/11/24● Data Quality : bad
Shen-Yi-Can-Mou
TMall’s Analytic Platform
Order # of products
1 3
2 2
3 1
Order Product ID
1 82275
1 84567
1 87632
2 82275
2 87632
Order List Item List
Compute date Product ID Product Name # of payments $ of payments
Data Preprocessing TMall Analytics
Order List Item List
Separate product names into one row
Find the product names and product id, if there has same name
but different id, change the id.
Join two list
If there has pId
Find the pId from pName
We cannot know which product it is
If there has pId
Aggregate sales data into daily and monthly data frame
Sheng-Yi-Can-Mou Data
Merge into the data frame
Forecasting data
No
Yes
Yes
No
We have 1314 row data don’t have pId, so we ignore it.
Method: Forecast Monthly Data
P82275 monthly forecast error (RMSE)
Model Training error Test error
sNaive 336.8237 93.0430
regression 129.15176 76.34134
arima 289.2212 222.6107
Ensemble 251.7322 130.665
Method: Forecast Daily Data
P82275 daily forecast error (RMSE)
Training error Test error
sNaive 11.582822 2.632218
ets 8.317506 2.138090
arima 8.317509 2.138103
Ensemble 9.405945667 2.302803667
EvaluationMonthly (RMSE) Daily (RMSE)
sNaive(Benchmark)
121.4645 4.53995822
38.87665 1.50777218
ets131.9892 3.315389
174.2751 1.00362794
regression40.12895 5.535291
41.63898 3.6440104
arima102.0755 3.294651
99.33205 0.98115324
Ensemble(top 3 model)
87.88965 3.716666073
59.9492265 1.164184453Over-forecast!
Overfitting!
Recommendation1. sNaive for monthly, ets regression for daily.
2. Retain their data autonomy, and confirm the data quality additionally.
3. Compare the similarities.
4. Data long enough inventory management reaching lean production.
5. Forecasting + Domain Knowledge
Constraint1. Forecasting future clothing trend is hard when solely using the data.2. Short product life cycle.3. Data constraint
Monthly
Recommendations & Limitations