ETA & Rider Assignment Root-Cause Analysis
Where delivery performance is breaking down across 15,948 delivered orders, and how much of it is prediction, assignment, or first-mile execution.
Executive summary
Khedmah has both a prediction-calibration issue and an execution-tail issue.
The median order is delivered roughly 13.5 minutes earlier than predicted, and 46.11% arrive more than 15 minutes early. However, 22% of orders still miss ETA, and approximately 7.2% are more than 15 minutes late.
Rider-to-restaurant execution is a material contributor to the late tail.
29% of orders take more than 20 minutes after rider assignment simply to reach the restaurant; 1 in 9 takes more than 30 minutes. This time is consumed before the final customer-delivery leg even begins.
Distance alone is not the root cause.
2,448 orders — 15.35% of the entire delivery network — had a rider <6 km away but still took >20 minutes to reach the restaurant. 813 orders — 5.10% of the network — took >30 minutes. This is too large to treat as an edge case.
These near-but-slow assignments are materially linked to ETA failure.
The <6 km / >20 min group alone contributes 1,014 ETA misses, equal to 28.92% of all ETA misses. The <6 km / >30 min group contributes 531 ETA misses, equal to 15.15% of every ETA miss in the dataset.
Khedmah should avoid concluding that “nearest rider” is the solution.
The evidence instead points toward the need to determine which rider is genuinely able to reach the restaurant fastest — which requires rider state, ongoing orders, GPS freshness, road ETA, rider movement, offer/rejection history and alternative eligible riders.
1 · Analysis base
The analysis covers 15,948 delivered orders. Predicted ETA is available for 15,939 orders (99.94% of total orders). Only 9 orders (0.06%) do not contain a usable predicted ETA. Unless explicitly stated otherwise, all “% Overall” figures use the complete 15,948-order population as the denominator.
The purpose of this analysis is to identify where delivery performance is breaking down and to distinguish between ETA prediction/calibration issues, rider assignment issues, rider-to-restaurant execution issues, geographic/supply issues, rider-specific patterns, vehicle-type effects and potential rider-state/GPS issues.
2 · Delivery performance vs predicted ETA
| Metric | Orders | % of overall orders |
|---|---|---|
| Total delivered orders | 15,948 | 100.00% |
| Predicted ETA available | 15,939 | 99.94% |
| Delivered before predicted ETA | 12,432 | 77.95% |
| Delivered after predicted ETA | 3,506 | 21.98% |
| Delivered exactly at ETA | 1 | 0.01% |
| Predicted ETA unavailable | 9 | 0.06% |
The median order is delivered approximately 13.5 minutes earlier than the predicted ETA.
Khedmah does not appear to have a system-wide problem of consistently giving customers overly aggressive ETAs. Instead, the prediction appears relatively conservative for a large part of the network, while a smaller but material portion of orders substantially exceed prediction.
Almost 46.1% of all orders are delivered more than 15 minutes earlier than prediction. At the same time, 21.98% miss predicted ETA — and 1,147 orders, 7.19% of the entire network, are more than 15 minutes late.
Prediction calibration
A large portion of orders receive a prediction substantially longer than their actual delivery time.
Operational reliability
A smaller but important tail experiences significant delays that cause the prediction to fail. Simply adding more buffer to ETA would hide operational problems rather than solve them.
3 · The core anomaly: rider is close, but still takes too long
At network level, 4,627 orders (29.01%) take more than 20 minutes after rider assignment for the rider to reach the restaurant, and 1,783 orders (11.18%) take more than 30 minutes. The important question is how much of that is explained by distance.
More than half of the >20-minute first-mile problems cannot simply be explained by the rider being far away. A rider who is apparently within 6 km but takes more than 30 minutes is almost 3× as likely as the network average to miss ETA.
4 · Distance vs rider-to-restaurant time
Across orders with usable distance and rider-arrival data, the Pearson correlation between rider distance at assignment and rider → restaurant time is only approximately 0.29. Distance clearly matters, but this is not a strong enough relationship to explain rider arrival performance on its own.
The most striking finding is the tail. Even among riders only 0–2 km away, P90 rider arrival is approximately 23 minutes; for riders only 4–6 km away, P90 is already 30.8 minutes.
Geographic distance is not the same thing as operational availability. A rider can be physically close but still be a poor candidate for assignment.
5 · Area-wise analysis
For area analysis it is important to distinguish volume — how many total ETA misses originate from the area — from severity, the percentage of that area's own orders that miss ETA. These are not the same problem.
| Service area | Orders | ETA misses | Miss rate in area | Rider >20 min |
|---|---|---|---|---|
| Al Khuwayr South | 1,688 (10.58%) | 276 (1.73%) | 16.4% | 405 (2.54%) |
| Salalah Central | 1,025 (6.43%) | 259 (1.62%) | 25.3% | 231 (1.45%) |
| Koudh2 | 812 (5.09%) | 213 (1.34%) | 26.2% | 288 (1.81%) |
| Mabelah1 | 826 (5.18%) | 195 (1.22%) | 23.6% | 301 (1.89%) |
| Amerat | 1,505 (9.44%) | 161 (1.01%) | 10.7% | 334 (2.09%) |
| Bawshar | 629 (3.94%) | 155 (0.97%) | 24.6% | 199 (1.25%) |
| Al Ghubrah North | 693 (4.35%) | 154 (0.97%) | 22.2% | 199 (1.25%) |
| Ruwi | 763 (4.78%) | 152 (0.95%) | 19.9% | 147 (0.92%) |
| Barka | 684 (4.29%) | 150 (0.94%) | 21.9% | 172 (1.08%) |
| Saadah South | 542 (3.40%) | 138 (0.87%) | 25.5% | 168 (1.05%) |
| Al Mawaleh South | 502 (3.15%) | 137 (0.86%) | 27.3% | 192 (1.20%) |
Largest contributors to ETA misses, ranked by absolute misses. Network ETA miss rate: 21.98%.
Al Khuwayr South contributes the largest absolute number of misses with 276, but its internal miss rate is only 16.4%, below the network average of 22.0%. That suggests a volume effect, rather than necessarily an area-performance problem. Conversely, some smaller areas have much higher failure rates.
These areas should be examined for:
- inadequate rider supply
- incorrect zone coverage
- rider concentration
- vehicle mix
- assignment strategy
- time-of-day supply imbalance
6 · Rider-wise analysis
Two metrics are required: absolute miss contribution — how many network misses came from the rider — and the rider's own ETA miss rate. The second is particularly important.
| Rider | Orders | ETA misses | Rider's miss rate | Rider >30 min |
|---|---|---|---|---|
| Tassawar Hussain | 429 (2.69%) | 132 (0.83%) | 30.8% | 87 (0.55%) |
| Muhammad Yaseen | 417 (2.61%) | 130 (0.82%) | 31.2% | 92 (0.58%) |
| Abdo Khaled Mohammed Ahmed Hamid | 304 (1.91%) | 127 (0.80%) | 41.8% | 105 (0.66%) |
| DAWOOD RAMADHAN | 300 (1.88%) | 115 (0.72%) | 38.3% | 74 (0.46%) |
| Ahmed Raza | 449 (2.82%) | 112 (0.70%) | 24.9% | 85 (0.53%) |
| Umer Rehman | 329 (2.06%) | 89 (0.56%) | 27.1% | 63 (0.40%) |
| Badar Munir | 516 (3.24%) | 87 (0.55%) | 16.9% | 41 (0.26%) |
| Aamer Hafeez | 337 (2.11%) | 87 (0.55%) | 25.8% | 49 (0.31%) |
| Mohammed Mamunur | 261 (1.64%) | 80 (0.50%) | 30.7% | 34 (0.21%) |
| Abdur Rahman | 216 (1.35%) | 78 (0.49%) | 36.1% | 40 (0.25%) |
| Dilawar Hussain | 246 (1.54%) | 72 (0.45%) | 29.3% | 26 (0.16%) |
| Usama Younas | 368 (2.31%) | 69 (0.43%) | 18.8% | 23 (0.14%) |
| MUSLIM | 188 (1.18%) | 69 (0.43%) | 36.7% | 44 (0.28%) |
| KHALIL | 318 (1.99%) | 65 (0.41%) | 20.4% | 48 (0.30%) |
| IMAM HOSSAIN | 210 (1.32%) | 65 (0.41%) | 31.0% | 18 (0.11%) |
These 15 riders handled approximately 30.65% of all orders but contributed 39.28% of all ETA misses.
The network ETA miss rate is 21.98%. Several riders are therefore operating at roughly 1.5×–2× the network miss rate. This should not immediately be interpreted as rider negligence. Possible explanations include rider behaviour, rider reliability, zones assigned to those riders, shift timings, vehicle type, long-distance assignments, concurrent or previous orders, and rider availability-state accuracy.
A rider-adjusted analysis controlling for area, distance and time of day would be required before attributing causality.
7 · Vehicle-type analysis
The dataset identifies vehicle types only as Vehicle Type 3 and Vehicle Type 4. The business mapping should be confirmed before management conclusions are drawn.
Vehicle Type 3 stands out materially, and warrants specific investigation once the underlying vehicle definition is confirmed.
8 · Root-cause interpretation
ETA calibration
The prediction is conservative for a significant proportion of the network: 77.95% arrive before predicted ETA, 46.11% more than 15 minutes early, median delivery approximately 13.5 minutes early. This suggests an opportunity to improve calibration and potentially give customers a more competitive promise — but it should be addressed after, or separately from, the operational tail.
Rider first-mile execution
A major operational concern: 29.01% take >20 minutes rider → restaurant and 11.18% take >30 minutes. Long first-mile execution materially increases the likelihood of ETA miss.
Proximity is not explaining the problem — arguably the most important root-cause finding
52.91% of >20-minute rider journeys occur at <6 km, and 45.60% of >30-minute journeys occur at <6 km. Simply reducing assignment radius or always choosing the geographically closest rider will not solve the complete problem. Potential causes requiring validation include:
Geographic supply imbalance
Certain areas have substantially higher miss rates than the 21.98% network average. This could reflect insufficient supply, rider deployment mismatch, rider type/vehicle concentration, area geometry and traffic, or assignment constraints. It should be investigated on an hour × area basis rather than area alone.
Rider / vehicle effects
There is meaningful variation across riders and vehicle types. However, rider performance should be normalized for area, order distance, shift, hour, vehicle, rider assignment distance and ongoing-order status before concluding that the rider themselves is the cause.
9 · Recommended next diagnostic
The next analysis should focus specifically on the 2,448 near-but-slow orders. For every one of these orders, reconstruct:
- Selected rider
- Rider state at assignment
- Previous/current active order
- GPS timestamp and location freshness
- Actual road ETA to restaurant
- Time rider started moving after assignment
- Rider offer/acceptance timing
- Other eligible riders available at the same moment
- Distance/ETA of those alternative riders
- Why the selected rider was preferred
This would allow Khedmah to classify the issue into:
Supply problem
No better rider existed.
Assignment logic problem
A materially faster rider existed but was not selected.
Rider-state problem
The rider appeared available but was actually occupied.
Acceptance problem
Closer riders rejected or timed out.
GPS / data problem
The system was using an inaccurate rider location.
Rider behaviour problem
The selected rider did not move toward the restaurant as expected.
The current data has moved the analysis beyond “Are riders being assigned too far away?” The much more important question is now:
Why do thousands of riders who are already geographically close to the restaurant still take 20–30+ minutes to reach it?
Answering that question is likely to expose the most actionable root causes in Khedmah's current rider-assignment and delivery operation.
Internal — Khedmah management only · Analysis period 5 Aug – 3 Sep 2026 · 15,948 delivered orders