Final Report: Olist

Exploration

1. Sales

---

LTM %>% 
  ggplot(aes(x=shipping_limit_date))+
  geom_histogram(binwidth=1,fill="white",color="navy")+
  geom_hline(yintercept = nrow(LTM)/365,color="darkred")+
  theme_minimal()+
  labs(x="Shipping Date",y="Count")

• The histogram above shows daily order shipping. The mean shipment is 281 order a day, and we notice interesting micro events such as Christmas period and another a slump in June 2018 for unknown reasons. The weekends are easily identifiable by the regular depressions. Noticeably, there is no black friday/thanksgiving sales. Manager could try to exploit this event in future years.

---

ggmap(Brazil_terrain,)+ 
  geom_point(data=LTM,aes(x= customer_long, y=customer_lat),alpha=0.3,size=0.1,
             color="navy")

## Warning: Removed 17 rows containing missing values (geom_point).

• On the scatter plot above each blue dot represents a postal code from where at least one customer has ordered. As can be seen we have a greater population in the southern part of Brazil than in the northern part, while the western regions, belonging to the Amazon rainforest are relatively desert.

---

• The graph above shows the demographic density per regions in Brazil. It seems as if Olist customers are not proportionally distributed with respect to the population repartition in brazil, and is heavily skewed toward the south.

• The map above depicts mean income per capita per region, a way of vizualizing socio-economic disparities throughout the country. We observe similar distribution patterns, with the richer people in Brazil living in the south and agglomeration of Rio and Sao Paulo. This map help us shed more light on the distribution of Olist customer. Intuitive we expect Olist customers to be upper-middle class population of brazil, a still developing country, and therefore most of them should live in the southern regions.

---

ggmap(Brazil_terrain,)+ #XX
  geom_density_2d(data=LTM,aes(x= customer_long, y=customer_lat),color="darkred")

## Warning: Removed 17 rows containing non-finite values (stat_density2d).

• This map again pictures olist customers but instead of scatter points, which are often inapropriate when dealing with thousands of overlaping points, we use a 2d density plot. The foremost inner circles indicate the highest densities of customers. It is even more obvious here that the bulk of Olist customer are concentrated in the region of Rio and Sao Paulo.

  ggmap(Sao,)+ #Density x customer 35/36
    geom_density_2d(data=LTM%>%group_by(customer_id)%>%
                 summarise(sum=sum(total_value),lon=mean(customer_long),
                                   lat=mean(customer_lat))%>%filter(sum>500/0.3),aes(
                                     x=lon,y=lat))+
    geom_density2d(data=LTM%>%group_by(customer_id)%>%
                      summarise(sum=sum(total_value),lon=mean(customer_long),
                                lat=mean(customer_lat))%>%filter(sum<100/0.3),aes(
                                  x=lon,y=lat),color="red",size=0.1,alpha=0.8)

## Warning: Removed 417 rows containing non-finite values (stat_density2d).

## Warning: Removed 46659 rows containing non-finite values (stat_density2d).

• The map of Sao Paulo above shows in red the density of customer who buy less than 100$ on average and in blue the density of customers who buy more than 500 dollar on average. We can see that blue customer density is much more circular and increasing as we progress toward city centers.

ggmap(Brazil_terrain,)+ #32
  geom_density_2d(data=LTM%>%
                    filter(price>500/0.3),aes(y=customer_lat,x=customer_long,),
                  size=0.5,color="blue")+
  geom_density_2d(data=LTM%>%
                    filter(price<100/0.3),aes(y=customer_lat,x=customer_long),
                  size=0.5,color="red")+
  theme(legend.position = "none",axis.title.x = element_blank(),
        axis.title.y = element_blank())

## Warning: Removed 16 rows containing non-finite values (stat_density2d).

• At a national scale, we observe the opposite effect, with blue customer density spreading more than red customers. The underlying hypothesis we developped is that people who live in the country side or outside big cities will be more likely to affect a greater share of their budget to buying online.

---

2. Prices

LTM%>%
  group_by(Month)%>%
  summarise(AveragePrice=mean(total_price))%>%
  ggplot(aes(x=Month,y=AveragePrice*0.3))+
  geom_col(fill = "Navy",color="Navy",alpha=0.5)+
  theme_minimal()+
  labs(x="Shipping Month",y="Average Dollar Value")

• The bar chart above gives us the mean price of an an order purchase (a basket) on the olist platform, that value is roughly 40 dollars, about half of the 78 dollar mean order value at Amazon. Considering the Brazilian income, on average 5 times smaller than in the US, we think that Olist mean order is relatively high, although to draw a more usefull comparison we should access the income of olist customers, which is arguably higher than the median income.

LTM%>%group_by(customer_id)%>%
  summarise(n=n()) %>% .$n %>% summary()

Min.	1st Qu.	Median	Mean	3rd Qu.	Max.
1	1	1	1.189167	1	34

• However on average, the Olist customer buys 1.18 package a year, a very low number if compared to Amazon customers, who’s 40% of customers buy once a week.This low turnover should be investigated by managers, for there is evidently a lot of margin for improvement.

---

LTM%>%
  filter(total_value<1000&total_value>0)%>%
  ggplot(aes(x=total_value))+
  geom_density(color="navy",fill="Navy",alpha=0.5)+
  theme_minimal()+
  labs(x="Value of Order",y="Density")

• As can be seen on the density chart above, the bulk of orders are distributed toward 40$ (here in local currency). That pareto-like distribution also present a long tail effect, as can be observed on the table below.

LTM%>%
  mutate(BinValue = cut(total_value,c(0,50,100,200,400,800,1600,Inf)))%>%
  group_by(BinValue)%>%
  summarise(VolumeSales = sum(total_value), Orders = n())%>%
  mutate(per_volume=VolumeSales/sum(VolumeSales),per_Orders=Orders/sum(Orders))

BinValue	VolumeSales	Orders	per_volume	per_Orders
(0,50]	556182.6	15099	0.0394860	0.1724122
(50,100]	1991768.5	27141	0.1414050	0.3099172
(100,200]	3917057.0	27570	0.2780903	0.3148159
(200,400]	3319098.7	12280	0.2356384	0.1402227
(400,800]	2076930.0	3806	0.1474510	0.0434599
(800,1.6e+03]	1445121.6	1350	0.1025960	0.0154154
(1.6e+03,Inf]	779399.7	329	0.0553332	0.0037568

---

3. Payments

LTM %>% #(6)
  group_by(payment_type)%>%
  summarise(vol=sum(total_value))%>%
  mutate(per=vol/sum(vol))%>%
  ggplot(aes(x="",group=reorder(payment_type,-vol),y=per,fill=payment_type))+
  geom_bar(color="black",stat="identity",alpha=0.8)+
  theme_minimal()+
  scale_fill_brewer(palette="Paired",name="Payment Type")+
  labs(x="Payment Type",y="% of Sales")

• Credit card and boleto account for over 90% of the sales. Boleto is a Brazilian method of payment that requires the user to go to the nearest bank to make a money transfer. Although time consuming, this method of payment is very popular because it is deemed more secure than using credit cards. Brazil is in fact notorious for its rate of credit card fraud, higher than in most countries.

LTM %>% 
  ggplot(aes(x=log(total_value),color=payment_type))+
  geom_density(size=1)+
  theme_minimal()+
  scale_color_brewer(palette="Spectral",name="Payment Type")

• Credit card payment distribution has a slight shift toward higher value payments, a phenomenon that can be analyzed more precisely in the table below.

LTM%>% group_by(payment_type)%>%
  summarise(n=n(),avg_spend=mean(total_value))

payment_type	n	avg_spend
boleto	16890	150.3465
credit_card	65423	165.4633
debit_card	1419	143.0357
voucher	3843	134.8241

t.test(LTM%>%filter(payment_type=="boleto")%>%.$total_value,
       LTM%>%filter(payment_type=="credit_card")%>%.$total_value)

## 
##  Welch Two Sample t-test
## 
## data:  LTM %>% filter(payment_type == "boleto") %>% .$total_value and LTM %>% filter(payment_type == "credit_card") %>% .$total_value
## t = -7.6851, df = 26753, p-value = 1.581e-14
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
##  -18.97228 -11.26133
## sample estimates:
## mean of x mean of y 
##  150.3465  165.4633

• As shown above, payments made with boleto are in general lower than payments made with credit card. This observation could prompt the hypothesis that endemic fraud in Brazil negatively and directly impacts Olist business through smaller sales. Proving such causal effect would require different sets of data.

---

4. Installments

• Olist customers who pay with credit card have the possibility to pay with many installments (probably the reason why credit card on average pay for higher value baskets). Unsurprisingly, the number of installments tends to have a positive relationship with the overall value of the items bought.

LTM %>% filter(payment_installments<=10,payment_installments>0)%>%
  ggplot(aes(x=as.factor(payment_installments)))+
  geom_bar(fill="Navy",color="Navy",alpha=0.5)+
  theme_minimal()+
  labs(x="Payment Installments",y="count")

LTM %>% 
  ggplot(aes(x=log(total_value), 
             group=cut(payment_installments,c(-Inf,2,8,+Inf)),
             color=cut(payment_installments,c(-Inf,2,8,+Inf))))+
  geom_density()+
  theme_minimal()+
  scale_color_discrete(name="Groups of Installments")

LTM %>% group_by(cut(payment_installments,c(-Inf,2,8,+Inf)))%>%
  summarise(avg_value = mean(total_value),n=n())

cut(payment_installments, c(-Inf, 2, 8, +Inf))	avg_value	n
(-Inf,2]	126.4004	54464
(2,8]	187.5936	27649
(8, Inf]	368.8228	5462

• What is more surprising however is the frequency with which customer use payment facilities. Half of customers use these and 60% of the overall sales volume is financed through payment facilities. Thus it seems that these options of payment are very important for Olist. Managers should perhaps contemplate encourage more use of payment facilities, after considering the opportunity cost for the company or whether consumer who use many installments of payment tend to run away with the money.

LTM %>% group_by(cut(payment_installments,c(-Inf,1,Inf)))%>%
  summarise(sm = sum(total_value),n=n()) %>%
  mutate(prop_n = n/sum(n),prop_val=sm/sum(sm))

cut(payment_installments, c(-Inf, 1, Inf))	sm	n	prop_n	prop_val
(-Inf,1]	5536983	43997	0.5023922	0.3930965
(1, Inf]	8548575	43578	0.4976078	0.6069035

---

5. Categories

• The bar chart below shows cumulative percentage of decreasing orders of product categories. We see for example that 75% of sales are concentrated among the top 15 categories of products. The second graph below shows those top 15 categories. It would be interesting for managers to study the sales correlation between those product categories and see whether any diversification gain can be achieved by growing other product segments, for example to protect themselves in case of economic shock. Other variables such as own-price elasticity or sales fees should be taken into account by managers when reflecting on how to change the product mix.

LTM%>% #(12)
  group_by(product_category_name_english)%>%
  summarise(vol=sum(total_value))%>%
  mutate(rank=rank(-vol))%>%
  arrange(rank)%>%
  mutate(cum=cumsum(vol))%>%
  mutate(cumper=cum/max(cum))%>%
  ggplot(aes(x=as.factor(rank),y=cumper,fill=rank))+
  geom_col(alpha=0.8,color="white",size=0.001)+
  theme_minimal()+
  geom_vline(xintercept = 3,color="darkred")+
  geom_vline(xintercept = 7,color="darkred")+
  geom_vline(xintercept = 14,color="darkred")+
  labs(x="Item Rank",y="Cumulative % of Sales")+
  theme(legend.position = "none",
        axis.text.x = element_text(size=4))+
  scale_fill_distiller(palette ="Spectral")

LTM %>%
  group_by(product_category_name_english)%>%
  summarise(vol=sum(total_value))%>%
  mutate(ShareVol=vol/sum(vol))%>%
  filter(vol>90000/0.3)%>%
  ggplot(aes(x=reorder(product_category_name_english,vol),y=ShareVol))+
  geom_col(aes(fill=vol),alpha=0.9,color="black")+
  theme_minimal()+
  labs(x="",y="% Total Sales")+
  theme(axis.text.x = element_text(angle = 90, hjust = 1),
        legend.position="none")+
  coord_flip()+
  scale_y_continuous(labels=percent)

• The colorful chart below shows the top 5 categories of product for each months of the year. Interestingly there is a lot of turnover and few categories stay at the top throughout the year. Only bed_bath_table and health_beauty seem to be year-round best sellers. Other categories such as housewares or computers_accesories seem more seasonal. Finally, some product categories make very punctual incursion such as computers (beginning of the school year probably).

LTM %>%
  group_by(Month,product_category_name_english)%>%
  summarise(vol=sum(total_value))%>%
  group_by(Month)%>%
  filter(vol>sort(vol,T)[6])%>%
  group_by(Month)%>%
  mutate(rank=rank(-vol))%>%
  ggplot(aes(x=Month,y=rank,fill=product_category_name_english))+
  geom_tile(colour="black")+
  theme_minimal()+
  labs(x="Month",y="Rank(1=best)")+
  theme(legend.title=element_blank(),
        legend.text=element_text(size=15))

---

5. Freight

LTM %>%
  group_by(product_category_name_english)%>%
  summarise(volume=sum(total_value),avg=mean(freight_value),wh=mean(product_weight_g))%>%
  filter(avg>(7/0.3))%>%
  ggplot(aes(x=reorder(product_category_name_english,avg),y=avg*0.3))+
  geom_col(aes(fill=volume<90000/0.3),color="black",alpha=0.8)+
  theme_minimal()+
  labs(x="",y="Average $ Delivery Price")+
  theme(axis.text.x = element_text(angle = 90, hjust = 1),
        legend.position="none")+
  coord_flip()+
  scale_fill_brewer(palette="Set1")

LTM %>%
  group_by(product_category_name_english)%>%
  summarise(ratio=sum(total_freight)/sum(total_value),vol=sum(total_value))%>%
  arrange(desc(ratio))%>%
  filter(ratio>0.18)%>%
  ggplot(aes(x=reorder(product_category_name_english,ratio),y=ratio))+
  geom_col(aes(fill=vol<90000/0.3),color="black")+
  coord_flip()+
  theme_minimal()+
  labs(x="",y="DeliveryFee/OverallValue")+
  scale_y_continuous(labels = percent)+
  scale_fill_brewer(palette="Set1", name="Top 15 Product",labels=c("TRUE","FALSE"))

• The bar charts above shows the delivery fee per product categories and the categories with the highest delivery to order value ratio. A first observation is that delivery fees seem quite expensive compared to the prices we know in developed economies, manager should perhaps see whether any efficiency can be achieved on that aspect, or see how hiding delivery fees in the product price would affect sales. Also, noticeable some product categories have prohibitively expensive delivery fee ratios. Among those are some of the best sellers product categories. Managers should investigate whether this ratio could be reduced any how, to increase margins or overall sales.

---

6. Sellers

LTM%>%  #17
  group_by(seller_id)%>%
  summarise(vol = sum(total_value))%>%
  mutate(per=vol/sum(vol))%>%
  arrange(desc(vol))%>%
  mutate(cum=cumsum(per),r=rank(-per,ties.method = "random"))%>%
  filter(r<2000)%>%
  ggplot(aes(y=cum,x=r))+
  geom_col(fill="navy",color="Navy",alpha=0.1)+
  geom_vline(xintercept = 27,color="darkred")+
  geom_vline(xintercept = 125,color="darkred")+
  geom_vline(xintercept = 420,color="darkred")+
  theme_minimal()+
  labs(x="Rank",y="Cumulative percentage of Sales")

• This cumulative distribution of sales in terms of individual sellers again results in a nicely unbalanced distribution, with 420 sellers (2566 in total) controlling 75% of the sales. Curiously, the proportions of that distribution is very much similar to that of wealth distribution in the world. Surely not a coincidental phenomenon.

ggmap(Brazil_terrain,)+ #18
  geom_point(data=LTM%>%
               group_by(seller_id)%>%
               summarise(vol = sum(total_value),lat=mean(seller_lat),lon=mean(seller_long))%>%
               mutate(per=vol/sum(vol,na.rm=TRUE))%>%
               arrange(desc(vol))%>%
               mutate(cum=cumsum(per),r=rank(-per,ties.method = "random"))%>%
               filter(r<125),aes(x=lon,y=lat,size=vol,color=r<27),
             alpha=0.5)+
  theme(legend.position = "none",axis.title.x = element_blank(),
        axis.title.y = element_blank())

• On the map above we see the location of the main sellers, in blue the top 27 (25%) and in red the top 127 (50%). Most are concentrated in the southern region where most of the business is being made. One outlier is the third biggest seller, located in the region of Salvador.

---

7. Delivery Process

LTM %>% # 19 
  ggplot(aes(x=clicktodoor))+
  geom_histogram(bins=60,fill="navy",color="Black",alpha=0.8)+
  xlim(0,60)+
  theme_minimal()+
  labs(y="Click-to-Reception Days")

## Warning: Removed 255 rows containing non-finite values (stat_bin).

## Warning: Removed 2 rows containing missing values (geom_bar).

• half of the orders arrives in 10 days to destination after the customer first clicked to purchase. This is much higher Amazon which orders on average arrive in only 3.28 days. This is surely a kpi which Olist should strive to improve, but this is nonetheless not a main point of preoccupation given the relative youth of the company. 3 years ago, Amazon package took on average 6 days to come to destination.

• The interesting distribution below, of orders as a function of distance traveled, is not smooth at all. The first two peaks can be explained in the following manner: Rio and Sao Paulo are the two main centers of economic activity and are separated by roughly 400 km, the distance between the peaks. The massive amount of bilateral trade between those cities, who trade either to their agglomeration or the agglomeration of the other cities, cause this pattern of booms and busts. The other peaks should also be the result of a similar phenomenon at play.

LTM %>% #20
  filter(distance<3000)%>%
  ggplot(aes(x=distance))+
  geom_density(fill="darkred",alpha=0.3)+
  geom_histogram(bins=100,fill="Navy",color="grey50",aes(y=..density..))+
  theme_minimal()

• On median, a package travels 427 Km, which, if divided by the median time taken by a package to reach its destination, gives 42km a day, which does not seem like a lot. Another key performance indicator that managers should focus on improving.

• The process of delivering a package is divided in 3 parts:

1. Customer click on purchase and waits for its payment to be approved
2. Once the payment is approved, the seller has to pack the order and dispatch it to carrier
3. the carrier then takes the order and delivers it to the customer

• Below is a graph that represents how long it takes for an order to be approved on average depending on the payment means. It could be infered again, that Boleto payment type are a brake to consumption given that it systematically increases the approval type and therefore affects customer who need their package quickly or buy compulsively.

LTM%>% #21
  group_by(payment_type)%>%
  summarise(mn=mean(approval_time),mnn=mean(review_score))%>%
  ggplot(aes(x=payment_type,y=mn))+
  geom_col(fill="Navy",alpha=0.8,color="black")+
  theme_minimal()+
  labs(x="Payment Type",y="Payment Approval Time in Days")

• Below is a line chart of the number of days betwen the moment the carrier takes charge of the parcel and delivers it, as a function of the distance. Naturally we observe a positive relationship. The mean time taken by the carrier is 9 days.

LTM %>% #24
  filter(transit_logistic<90,distance<4000)%>%
  ggplot(aes(x=distance,y=transit_logistic))+
  geom_smooth(aes(x=distance,y=transit_logistic))+
  theme_minimal()+
  labs(y="Carrier Transit")

## `geom_smooth()` using method = 'gam' and formula 'y ~ s(x, bs = "cs")'

LTM %>% #23
  group_by(Sendoff_time)%>%
  summarise(cn=n(),avg=mean(review_score))%>%
  filter(Sendoff_time>=0& Sendoff_time<35)%>%
  ggplot(aes(x=Sendoff_time,y=avg))+
  geom_col(fill="Navy",color="black",alpha=0.7)+
  geom_smooth()+
  theme_minimal()+
  labs(x="Sendoff Time",y="Review Score")

## `geom_smooth()` using method = 'loess' and formula 'y ~ x'

• The graph above is supremely insightful and puzzling. It dispplays the average rating as a function of the number days taken by the seller to dispatch the order to the carrier,that is 9 days on average before the customer receives the package. As evidenced, that relationship is almost perfectly linear, and the mean difference is significant at least for the 6 first steps. Then the test looses significance as the sample sizes become too small.

LTM %>% #25
  filter(transit_logistic<90,distance<4000,price>10/0.3) %>%
  group_by(product_category_name_english)%>%
  mutate(bigpay=ratio_delivery>=1*median(ratio_delivery))%>%
  group_by(product_category_name_english,bigpay)%>%
  summarise(avg=mean(transit_logistic))%>%
  ggplot(aes(x=reorder(product_category_name_english,-avg),
             y=avg,
             fill=bigpay))+
  theme_minimal()+
  geom_col(position="identity",alpha=0.5)+
  theme(axis.text.x = element_blank())+
  labs(x="Categories",y="Average Transit Days")+
  scale_fill_brewer(palette = "Set1")

• The chart above has product categories on the X axis and average transit days on the Y axis. In red are the average transit days of the order for which the customer paid lower than the median delivery price and in blue, those who paid higher. Interestingly, those customer who pay more than the median delivery fee tend to receive their package later, almost systematically. The differences are significant when the sample size is large enough.

LTM$lag <- LTM$expectedtransit-LTM$transit

LTM%>% #26
  ggplot()+
  geom_smooth(aes(x=distance,y=expectedtransit))+
  geom_smooth(aes(x=distance,y=transit),color="red")+
  theme_minimal()+labs(y="Red=actual/Blue=Expected")

## `geom_smooth()` using method = 'gam' and formula 'y ~ s(x, bs = "cs")'
## `geom_smooth()` using method = 'gam' and formula 'y ~ s(x, bs = "cs")'

## Warning: Removed 7 rows containing non-finite values (stat_smooth).

• The graph above shows in blue the estimated days before delivery and in red the actual days. Strikingly, actual delivery takes 10 days less than predicted delivery with a very similar curve pattern (until a point). Secondly, Predicted delivery starts at 2 weeks for a distance of 0 kilometers, and goes up to 3 weeks for 200km deliveries. Finally, the line for predicted delivery is monotonously increasing as while the red has a concave shape. The concave shape could be explained by the fact that, as distance increases, or if the customer lives on an island, the carrier may shift to other means of transportation, such as planes, which increases the average travel speed of a parcel. All these element point out toward an obsolete software to estimate delivery times. The managers should investigate this in priority. It is not a strong statement to make that seeing two weeks estimated delivery time (at minimum) would deter many buyers.

LTM %>% #27
  filter(transit<60)%>%
  ggplot(position="identity")+
  geom_density(aes(x=transit),fill="Navy",alpha=0.5)+
  geom_density(aes(x=expectedtransit),fill="red",alpha=0.5)+
  theme_minimal()+
  ggtitle("Blue = Actual , Red = Predicted")+
  labs(x="Transit Days",y="Density")

t.test(LTM$transit,LTM$expectedtransit)

## 
##  Welch Two Sample t-test
## 
## data:  LTM$transit and LTM$expectedtransit
## t = -263.74, df = 174335, p-value < 2.2e-16
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
##  -11.49431 -11.32473
## sample estimates:
## mean of x mean of y 
##  11.93916  23.34867

LTM %>% #28
  filter(lag>-30&lag<50)%>%
  ggplot(aes(x=as.integer(lag)))+
  geom_histogram(aes(y=..density..),bins=80,fill="navy",color="grey50")+
  geom_density()+
  theme_minimal()+
  geom_vline(color="darkred",xintercept = 0,size=1)

• Above is vizualized the lag days, which is the difference in day between the estimated delivery days and the actual delivery days. We notice that the lag is highly volatile (indicating once again an ineffective method for predicting delivery date) and also massively positive (mean =11.5). However, despite the very conservative estimated delivery date, 7.5% of order come late.

ggmap(Brazil_terrain,)+ #29
  geom_density_2d(data=LTM%>%
               filter(lag<0),aes(y=customer_lat,x=customer_long,size=-(as.integer(lag))),
             alpha=0.8,color="red")+
  geom_density_2d(data=LTM%>%
                    filter(lag>10),aes(y=customer_lat,x=customer_long,size=-(as.integer(lag))),
                  alpha=0.8,)+
  theme(legend.position = "none",axis.title.x = element_blank(),
        axis.title.y = element_blank())

## Warning: Removed 2 rows containing non-finite values (stat_density2d).

## Warning: Removed 7 rows containing non-finite values (stat_density2d).

• The map above shows us in red the parcels that have come late and in blue the parcels that have come more than 10 days before estimated. We notice that the late parcels are not located particularly far from densely populated region (they are not in the amazon rainforest!), thus the hypothesis that distance or complex travel routes may explain lags seems unwarranted.

---

8. Reviews

LTM %>% ggplot(aes(x=review_score ,y=stat(count)/sum(stat(count))))+
  geom_histogram(color="black",fill="navy",alpha=0.8)+
  theme_minimal()+
  labs(x="Review Score",y="Count")+
  scale_y_continuous(labels=percent)

## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.

• Reviews have a U shaped distribution, which intuitively makes sense because people are emotional beings and tend to make extreme statements, thus for example, when you want to express your frustration, you are unlikely to think long between giving a two or a one. It is usually all-out.

LTM %>% group_by(product_category_name_english)%>%
  summarise(mn=mean(review_score),vol=sum(total_value),cn=n())%>%filter(cn>500)%>%
  arrange(-mn)%>%
  ggplot(aes(x=reorder(product_category_name_english,mn),
             y=mn-3,
             fill=vol>90000/0.3))+
  geom_col(color="black",alpha=0.8)+coord_flip()+
  scale_fill_brewer(palette = "Set1",name="Top 15 Sales")+
  scale_y_continuous(labels = function(x){x+3})+
  labs(y="Rating",x="")+
  theme_minimal()

LTM %>% filter(lag>-50&lag<50)%>%ggplot(aes(x=lag))+
  geom_density(aes(color=as.factor(review_score)),size=0.5)+
  scale_color_discrete(name="Review Score")+
  theme_minimal()+
  labs(x="Day Lags",y="")+
  scale_color_brewer(palette="Spectral")

## Scale for 'colour' is already present. Adding another scale for
## 'colour', which will replace the existing scale.

• The graph above show us the distribution of lags per review given. Noticeably, 1 star reviews have anoticeable bump with late deliveries, strenghtening the reasoning that late deliveries storngly impacts customer review.

LTM$Timetoreview <- LTM$review_answer_timestamp-LTM$order_delivered_customer_date
LTM %>% filter(Timetoreview>-5&Timetoreview<10)%>% 
  ggplot(aes(x=Timetoreview,
             fill=as.factor(review_score),
             y=..density..))+
  geom_histogram(position="dodge")+
  theme_minimal()+
  scale_fill_brewer(palette="Spectral")+
  theme(legend.position="none")

## Don't know how to automatically pick scale for object of type difftime. Defaulting to continuous.

## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.

• Above, on the x axis is the number of days taken by customer to review the product after receiving the product. We can see in red the distribution of one star reviews, and see that many of these reviews are given even before delivery of the parcel!