Workforce Capacity Planner
Do three warehouse sites have enough workers for the expected workload, and if not, what is the cheapest safe fix?
This repo is a three part series on fulfillment operations:
- Forecast demand. How much work is coming? (
scripts/forecast_demand.py, see Technical details)
- Plan capacity. Do the sites have enough workers? (this page)
- Protect parcels on the middle mile. When trucks run late, which parcels are at risk and what is the cheapest fix? (middle_mile/) With the 6 worker plan, trucks cause 446 late parcels (37 from real 2023 snow and heavy rain, 365 on one tight Chicago to Indianapolis schedule); with 10 workers, 138, because understaffing leaves no time to absorb a late truck.
- Stock the shelves and reach the door. How much inventory should each site hold, and should stations run vans or use a parcel carrier? (inventory/, last_mile/, see Part 4 below)
The answer
Each modeled site covers at least 95% of its workload on most days, but none removes every shortfall hour across the recorded 2023 workload. The analysis found two staffing options:
| Option | Workers | Overtime | Stress cases passed | Modeled cost |
|---|
| Affordable base case | 6 | Up to 2 hours | 12 of 36 | $106,326 |
| Lowest risk | 10 | Up to 2 hours | 36 of 36 | $176,616 at most |
Choose the 6 worker plan when productivity is near 40 units per productive hour and attendance is stable. Choose the 10 worker plan when low productivity or two worker absence is possible. See DECISION_ANSWER.md for the site by site answer.
At real labor costs. These costs use the model's assumed $22 an hour. Real 2023 data from BLS puts a warehouse worker at about $33 an hour to the employer (pay plus benefits and payroll taxes), so both plans cost roughly half as much again, and the gap between them grows. Replayed on 25 four week windows of real 2023 demand, 6 workers with overtime and temps on call beat 10 workers by $41,239 every 4 weeks at real costs, against $27,200 at assumed costs, with the same on time service. See Checked against real data.
What this means for the business
In plain English. Dollar amounts use the model's assumed wages, so they are planning estimates, not money already saved.
| Takeaway | The number | What to do |
|---|
| Temps on call beat permanent hiring | About $27,200 saved every 4 weeks (39% cheaper, roughly $350,000 a year); at real 2023 labor costs from BLS, about $41,200. On real 2023 demand the lean plan still shipped 99.7% of orders on time, against 100% for 4 extra permanent workers. | Keep 6 workers, with overtime and temps on call for busy days. |
| A better forecast is not worth buying | A perfect forecast would save only about $260 every 4 weeks. Daily orders swing too much to predict well (the best model still misses a day by 60%). | Spend on flexibility, not forecasting tools. |
| Fix staffing before trucks | Fixing a risky truck costs about $30 per late parcel saved; hiring costs about $158. But 60% of late parcels already leave the warehouse late. | Staff the warehouse first, then fix the riskiest trucks. |
| Know the two warning signs | The lean plan survives only 12 of 36 tough weeks. It fails when workers are slower than usual or two people are out. | When either shows up, call in temps right away. |
Part 4: inventory and last mile
Inventory (inventory/build_inventory.py). Reorder points for every site, product and size are set on January to June 2023 order lines and tested on July to October, which they never saw. Order lines, products, sizes and ABC classes are real; supplier lead times (5 to 11 days), unit cost and holding cost are assumed.
| Policy | Orders waiting for stock | Average stock value |
|---|
| 80% target for every item | 41% | $565,000 |
| 95% target for every item | 23% | $1,079,000 |
| 99% target for every item | 15% | $1,899,000 |
| By class (A 99%, B 95%, C 90%) | 22% | $1,314,000 |
| 95%, supplier always on time | 23% | $1,080,000 |
| 95%, re-planned every month | 24% | $1,080,000 |
More stock helps with falling returns, and neither a perfectly reliable supplier nor monthly re-planning changes much. The lines that run short explain why: 61% are one order bigger than two weeks of that item's normal demand, 27% are items never sold before July, and 36% are slow sellers with no safety stock (a line can have more than one reason). For scale, a 95% target holds 1.42 months of sales in stock, close to the 1.32 months US retailers held in 2023 (FRED); a 99% target holds 2.49. So the cheaper fix is a quoted lead time for unusually large orders and stock for new items at launch, not more of everything. Higher targets for A items cost more than a flat 95% and delay about as many orders, because one missing slow seller holds up the whole order.
Last mile (last_mile/build_last_mile.py). Every 2023 order comes out of the middle mile with its delivery station and day, for each staffing plan. Each station's day is routed with OR-Tools (van capacity, a 10 hour shift, fewest vans first), then the cheapest mix of owned and contracted vans is compared with a parcel carrier. Customers sit at real 2023 Census tract centers inside each station's 45 minute drive area (OpenRouteService), picked with odds matching each tract's population (ACS 2022); which tract each customer gets, van day rates and the carrier price are assumed. Van fuel is priced at each week's real 2023 US diesel price (FRED).
| 6 worker plan | 10 worker plan |
|---|
| Running vans, cost per order | $33.01 | $32.64 |
| Parcel carrier, cost per order | $15.29 | $15.29 |
| Orders per van a day (break-even about 42) | 19.5 | 19.8 |
以上為 README 的開頭部分。其餘內容請閱讀原文。 檢視完整 README ↗
