Operations Management · Token-Economy Design · SF Express field data (2024.07 – present)

Steering a
Reward-Token Economy

Issuance, Pricing, and Budget Targeting in a Dual-Cost Incentive Currency

SF Express pays couriers in Fengdou, a closed-loop token earned through work and redeemed for either zero-cost virtual goods or budget-consuming real goods. The firm sets how many tokens to issue and what they cost — but not how couriers spend them — and must land real-goods spending close to an annual budget it loses if it underspends.

Second layer these levers are set by a forecasting-and-inventory algorithm and adjusted by human operators — a human–AI collaboration we also study, using the recorded recommendation-vs-override trace.

The Design Problem → The Human–AI Layer
3
Control levers
3
Competing goals
≈¥40M
Annual real-goods budget
2 yrs
Transaction-level panel

The Fengdou System

A closed corporate token economy with a defining feature: two redemption channels with heterogeneous marginal cost to the firm.

Earn through work

Tokens accrue per delivery job. Issuance volume is the "money supply" — a firm-controlled lever.

🎮

Virtual goods (cost = 0)

Skins, badges, status items — real utility to couriers, zero marginal cost to SF.

📦

Real goods (cost > 0)

Paper, food, daily necessities — every redemption draws on the annual budget.

The bite: the budget must be spent close to target — overspend is infeasible, and underspend triggers a budget cut next year (a use-it-or-lose-it ratchet). Levers are set before redemption is observed, and the firm cannot dictate the virtual-vs-real split. It can only steer.


The Design Problem

Three levers, three goals — and a near-clean mapping between them. The virtual-goods price is a free budget-steering valve.

Lever Primarily steers Mechanism
1 Token issuance Incentive power + float More tokens ⇒ stronger effort signal, larger outstanding liability.
2 Real-goods prices Budget burn rate Sets how fast real redemptions draw down the annual budget.
3 Virtual-goods prices Budget steering — the valve Diverts token demand toward the zero-cost channel at will.
Goal 1 · Incentivize

Maximize effort / retention value per yuan of budget.

Goal 2 · Hit the budget

Land real-goods spend near target — avoid overspend and the underspend ratchet.

Goal 3 · Stable float (secondary)

Keep outstanding Fengdou in a healthy band — avoid hoarding or collapse.

Illustrative. As virtual-good attractiveness λ rises, real-goods spend falls; the two-sided budget penalty is U-shaped with an interior optimum λ*.

Pressure-valve principle

Virtual goods absorb token demand at zero cost, so their relative price — not issuance — is the sharpest budget-steering instrument. Issuance is then freed to do its real job: motivation.

Budget targeting

Levers commit before stochastic redemption realizes, with a two-sided over/under penalty — a newsvendor-type critical-fractile policy, with the virtual good thinning the tails.

Dynamics

Multi-period: token accumulation as a Miller–Orr float band, and the use-it-or-lose-it ratchet making the firm's problem an MDP.


Second layer

The Human–AI Layer

A forecasting-and-inventory algorithm recommends the levers; human operators adjust them with private information the model lacks — the ratchet, HQ pressure, fairness. We observe every step.

Algorithm

1 · Forecast & optimize

Predicts redemption demand and the float; recommends issuance & prices.

Human

2 · Adjust with private info

Operators override using ratchet risk, HQ pressure, campaigns, fairness.

Policy

3 · Commit final levers

The implemented vector — which may differ from the recommendation.

Outcome

4 · Realized result

Budget adherence · incentive value · float — the override's score.

The data asset. Both the algorithmic recommendation and the human-set decision are recorded, so we can isolate where operators deviate and whether each deviation moved the firm toward its goals — separating information deviations (the human knows more) from complexity / discretion deviations (the human won't or can't follow exactly).


Research Questions

Two on the token economy itself; one on the human–AI layer.

Q1

Optimal levers

How to set issuance and dual-cost pricing to hit the budget target while maximizing incentive value — the pressure-valve / separation result.

Q2

Float & ratchet dynamics

How to manage the token supply over time — the float band and the multi-year use-it-or-lose-it ratchet that distorts the static optimum.

Q3 2nd layer

When should humans override?

Using the recommendation-vs-override trace: which operator deviations from the algorithmic levers add value, and how to make the recommendation adherence-aware.


Where It Sits in the Literature

Primarily a token-economy / incentive-currency paper; secondarily a human–AI collaboration paper.

Token economics · platform & crypto currencies

  • Tsoukalas & Falk — "Token-Weighted Crowdsourcing," Manag. Sci. 2020. Tokens as an incentive mechanism — closest to ours. link
  • Sockin & Xiong — "A Model of Cryptocurrencies," Manag. Sci. 2023. Utility-token valuation & supply. link
  • Cong, Li & Wang — "Tokenomics: Dynamic Adoption and Valuation," Rev. Financial Studies 2021. Dynamic token economy. link
  • Meng, Hao & Tan — "Freemium Pricing with Virtual Currency," Inf. Syst. Res. 2021. Pricing/issuing a platform currency. link

Reward points · loyalty · float

  • Stourm, Bradlow & Fader — "Stockpiling Points in Linear Loyalty Programs," J. Marketing Res. 2015. Why holders hoard a points currency. link
  • Breakage in loyalty programsInt. J. Res. Marketing 2024. Unredeemed-reward liability — the accounting twin of token float. link
  • Putting Teams into the Gig EconomyManag. Sci. 2022. Gamified workforce incentives at a ride-share platform. link

Human–AI collaboration · second layer

Special Issue: The Human–Algorithm ConnectionManag. Sci. 72(1), 2026.

Kesavan, Kushwaha & Steele — judgmental adjustments lift profit (+4.92%), Manag. Sci. 2025. link

Kesavan & Kushwaha — merchant overrides cut profit 5.77% on average, Manag. Sci. 2020 — value is conditional. link

Balakrishnan, Ferreira & Tong — collaboration with private information, Manag. Sci. 2026. link

Adherence-Aware Recommendations — "The Best Decisions Are Not the Best Advice," Manag. Sci. 2024. The template for Q3. link

Angelova, Dobbie & Yang — "Algorithmic Recommendations and Human Discretion," Rev. Econ. Stud. 2025. Identifies which overrides add value. link

The gap we fill: the token-economy canon is dominated by tradeable crypto tokens and consumer loyalty points. Fengdou is neither — a closed-loop, non-tradeable incentive currency with dual marginal cost (free virtual / budget-consuming real) and a two-sided budget target under a use-it-or-lose-it ratchet. The human–AI layer is a second contribution on top of that design.


Intended Contributions

Targeting a UTD-24 OR/OM outlet (Management Science / Operations Research).

01

A new problem class

Formalizes the closed-loop, dual-cost, budget-targeted incentive currency — distinct from tradeable crypto tokens and consumer loyalty points.

02

Pressure-valve & budget targeting

A separation result — issuance for incentives, relative virtual/real pricing for the budget — and a two-sided, newsvendor-type targeting policy.

03

Float & ratchet dynamics

Token-supply management as a float band, and how the multi-year use-it-or-lose-it ratchet distorts the static optimum.

04 2nd layer

Value of human overrides

Using the recommendation-vs-override trace, identifies which operator deviations from the algorithmic levers improve budget adherence and incentive value.


Data

Full-granularity SF Express panel — issuance, balances, prices, individual redemptions (and the algorithm-recommendation-vs-human-decision trace), July 2024 → present.

Approach

Analytical-normative token-economy model with theorems, calibrated on the panel; dynamic (multi-period) float and ratchet. Human–AI overrides as a second, empirically-identified layer.

Status

Active. Model formulation and calibration in progress; counterfactual policy experiments to follow.