Delivery experience root cause analysis
Reviews decide marketplace ranking and whether customers come back. I measured how much late delivery drives 1–2 star reviews, then traced each order through approval, seller handling and carrier transit to find which step to fix first.
- Data
- Real: Olist public dataset, 94,819 delivered orders, Jan 2017 – Aug 2018
- Tools
- Python, pandas, statsmodels
- Methods
- Process-stage decomposition, logistic regression, segmentation
- Question
- Which part of fulfilment should operations fix first?
A few days late changes everything
On-time orders get a 1–2 star review 9% of the time. One to three days late doubles that. At 4+ days late, the majority of reviews are 1–2 stars.
Lateness alone doesn't prove cause: late orders might also be bigger, pricier or longer-haul. So I fitted a logistic regression controlling for those factors.
Share of 1–2 star reviews by days late
Hover for average rating and order count
What raises the odds of a 1–2 star review
Odds ratios from logistic regression (log scale; 1 = no effect). Hover for 95% confidence intervals.
Lateness is the strongest driver
Holding everything else equal, a late order has 13.9× the odds of a bad review (95% CI 13.2–14.7).
The second-strongest factor is an order split across multiple sellers (5.0×). That usually means several parcels arriving on different days.
Shipping across state lines has no significant effect of its own (p = 0.18) once lateness is included. Distance matters only when it makes the order late.
Where the delay happens
Median days spent in each stage
Late vs on-time orders
Late orders spend 23.9 days with the carrier vs 6.9 for on-time orders. Seller handling slips too (3.0 vs 1.8 days), but carrier transit is where most of the time goes.
When the seller misses the handover deadline
Late rate by whether the seller reached the carrier before shipping_limit_date
Missing the handover deadline makes lateness 3.8× more likely. Still, 73% of late orders were handed over on time, so sellers aren't the main cause.
Late rate by month
% of orders delivered after the estimate, by purchase month
It's systemic and seasonal: 14% around Black Friday and 16–21% in Feb–Mar 2018, vs 3–5% in normal months. The sellers with the most late orders are just the biggest sellers (9.6% late vs 8.2% on average).
States with the highest late rate
Customer state, minimum 300 orders. Overall average is 8.2%.
The north-east (AL, MA, PI, CE, SE, BA) and Rio de Janeiro are late up to 3× as often as the average.
Estimates are very conservative: the median on-time order arrives 12.3 days early. When something goes wrong, the delay still overruns that buffer.
Recommendations
Plan carrier capacity for peaks
Pre-book capacity and add backup carriers before Black Friday and Q1, when late rates triple. Carrier transit is where late orders lose most of their time.
Alert on missed seller handover
Flag orders not handed over by
shipping_limit_dateand nudge the seller the same day. Late risk is 3.8× higher once the deadline passes.Contact customers before they complain
When an order crosses its estimated date, send an apology and a voucher. Target the 4+ days-late group, where most reviews turn negative.
Fix split shipments and estimates
Consolidate multi-seller orders or set expectations clearly. Recalibrate delivery estimates per route: 12 days of padding may cost conversions without protecting customers.