Arnav Chudiwale
Decision 2 · Circular supply chainAll decisions

Can our returns become revenue instead of waste?

The answer

Yes. Grade every returned phone, and open two refurbishment centers: one in Fresno, one in Reno.

Recovered per phone
$22 → $141
from a flat recycler price to graded resale
Three-year value
$2.6M
net present value, base case, after all running costs
Under stress
3 of 3
scenarios stay profitable; prices can fall 25% before it breaks even

Independent project · Modeled on public data · January to February 2026 · Python, Gurobi, ArcGIS Pro, Power BI

Fig. 1 · Today: every phone, one price
$22recovered per phone
Each dot is a returned phone. Grading sends it where it is worth the most.
I · feasible

Every phone, one price.

A retailer was throwing away most of the value in its returns, and every option on the table would have worked. The question was which one was worth the most.

  • 54 stores across California, at real store locations of a major electronics retailer, take back about 28,000 phones a year.
  • Every one goes to a bulk recycler for a flat $22, whatever its condition, on a cycle of 60 days or more.
  • A phone loses about $4 of resale value a week while it waits.
  • A quarter are like new and 45% only need repair. The recycler pays the same for them as for a broken screen.
  • Five candidate sites could host a refurbishment center, each able to process 20,000 to 25,000 phones a year. No single site can take all 28,140, so of the 31 possible networks, 26 can serve all 54 stores.
26feasible networks
Fig. 2 · Real store locations. Each frame is one possible network.
II · optimal

The closest sites were not the cheapest.

I asked the same network two different questions. They gave two different answers, and only one of them saves money.

Fig. 3 · Every line is one store's returns.
Ask the network
ClosestFremont + Ontario
Average trip
54 miles
Facility cost
$2.03M a year
CheapestFresno + Reno
Average trip
258 miles
Facility cost
$1.54M a year

Phones travel almost five times farther, and the network costs $490,000 less every year.

Why: a building costs about 1,200 times more than the trucks.

Running two centers, per year$1,540,000
Moving every phone to them, per year$1,251
barely a line

A phone weighs under half a pound, so moving it 100 miles costs under two cents. Choosing the cheapest buildings matters far more than choosing the closest ones.

Where the value comes from: grading.

Once phones are sorted by condition, each one goes where it is worth the most. Value recovered is net of processing and of the time each phone spends in transit.

$22$141per phone, on average
25%
45%
30%
Like newTested, wiped, resold
RepairableScreen or battery replaced, resold refurbished
PartsComponents harvested, the rest recycled
III · engineered

It still works when the market turns.

$2.25M of capital, $1.55M a year to run (the $1.54M of buildings plus freight and handling), and the $619,000 recycler income given up. Move the two things that matter most and watch the answer.

What a refurbished phone sells for, against the plan.
How many phones come back each year, against 28,140.
Three-year net present value, at 10%
$2.6M
Payback 14 monthsAdded value, year one $1.87M
Worth doing.
Conservative$0.9M

Prices 10% lower, no growth in returns, costs rising 4% a year.

Base case$2.6M

Plan prices, returns growing 3% a year, costs rising 2%.

Optimistic$4.1M

Prices 10% higher, returns growing 5% a year, costs rising 1%.

What actually moves the answer: prices and volume. Not freight.

How to read it: each bar is one input. I moved that input 20% below plan (grey) and 20% above plan (gold), kept everything else at the plan, and measured how far the three-year value moved. The longer the bar, the more the decision depends on getting that input right. A 20% swing in resale prices moves the value by about $2.1M; the same swing in freight moves it by under $1,000.

Input moved ±20%Change in three-year value
20% lower20% higherValue = three-year net present value at 10%
To
Chief Financial Officer
From
Arnav Chudiwale
On
What to do with 28,000 returned phones a year

Approve two centers. Then protect the price.

  1. Open Fresno and Reno. They are the cheapest network, not the closest, and the difference is $490,000 a year.
  2. Spend on grading and resale, not logistics. Freight barely moves the result; grading accuracy and resale channels move almost all of it.
  3. Lock in buyers before scaling. The case survives a 25% fall in resale prices. Price agreements keep it there.
  4. Track one number monthly: value recovered per phone against the $141 plan.
Arnav Chudiwale

What the model can't tell you.

  • Store locations are real; return volumes and the grade mix are estimates from public return-rate and condition data, not a retailer's records.
  • Distances are straight-line, multiplied by 1.3 for roads. No seasonality, and returns are spread evenly across stores.
  • Grading is assumed 95% accurate, and facility capacity scales in a straight line.
  • An earlier version of this model left out running costs. The figures on this page are corrected.

For technical readers.

Models, data and tools

The decision. Which of five candidate sites to open, and which site each store sends its returns to. This is a facility-location model: a mixed-integer program solved in Gurobi.

Sets
i = the 54 stores; j = the 5 candidate sites (Fremont, Ontario, San Diego, Fresno, Reno)
Given
si = phones returned at store i each year (28,140 in total)
dij = road miles from store i to site j (straight-line × 1.3)
t = freight per phone per mile ($0.000172)
h = handling per phone ($0.35)
fj = yearly cost of running site j ($720,000 to $1.05M)
Kj = phones site j can process a year (20,000 to 25,000)
p = number of sites to open
Decide
yj = 1 if site j opens, else 0
xij = 1 if store i sends its returns to site j, else 0

Objective · lowest yearly costminimize   Σj fj yj  +  Σi Σj si ( t dij + h ) xij
ConstraintsΣj xij = 1   for every store  each store uses exactly one site
xij ≤ yj   for every store and site  only open sites receive phones
Σi si xij ≤ Kj yj   for every site  no site over capacity
Σj yj = p  open exactly p sites

How it was used. Solved for p = 2 to 5 (one site cannot process all 28,140 phones); two sites give the lowest total. The distance benchmark keeps the same constraints but minimizes Σ si dij xij (phone-miles) and picks Fremont + Ontario. As a first cut, ArcGIS Pro Network Analyst location-allocation, which ignores capacity, picked Ontario alone.

The business case. NPV = −C + Σt=1..3 CFt / 1.1t, where each year's added cash CFt = (R − B)(1 + g)t − O(1 + i)t. R = value recovered ($3.97M a year), B = recycler income given up ($619,000), O = running cost of the network ($1.55M), C = capital ($2.25M), g = growth in returns (3%), i = cost inflation (2%).

Data. 54 real store locations; site costs from California industrial rents; Haversine distances over 270 store-site pairs; freight from 2024 California less-than-truckload rates; Power BI dashboards.

Code on GitHub →