Operations Management · Incentive-Token Systems · Predict-then-Optimize

Steering an Incentive-Token System:
Issuance, Pricing, and Budget Targeting

A predict-then-optimize framework for a closed-loop, dual-cost, budget-targeted worker incentive currency — evidence from SF Express's Fengdou token program (400k+ couriers).
Xia Cai, Qingwei Zeng, Chao Zhou (SF Technology)  ·  Yue Zhao (HSBC Business School, Peking University)
Research Presentation  ·  ~10 minutes
Research collaboration with SF Express  ·  figures rounded / illustrative  ·  data de-identified.
FengdouSteering an Incentive-Token System

Problem Setting · 1 / 6

The Fengdou Incentive-Token System

SF Express pays more than 400,000 couriers in Fengdou — a non-tradeable token earned by completing delivery and service tasks, and spent in an internal mall.

🚚Work
tasks completed
→
🪙Issuance
firm sets the rate
→
🏦Token float
outstanding balance
→
🎮Virtual goods
cost ≈ 0
📦Real goods
draw on budget

Firm = bank + merchant

The firm sets the token supply and every price — but does not choose how workers spend. It can only steer.

Dual marginal cost

Virtual redemption is free to the firm; real redemption consumes a real annual budget (≈ ¥40M).

The bite

Real-goods spending must land close to the budget: overspend is infeasible, and underspend ratchets next year's budget down (use-it-or-lose-it). Levers are set before redemption is observed.

Problem SettingThe Fengdou Incentive-Token System

Problem Setting · 2 / 6

A New Problem Class

Fengdou is neither a tradeable crypto/utility token nor a consumer loyalty point — it is a third kind of currency:

📈

Crypto / utility token

  • Held by investors
  • Tradeable, market-priced
  • No firm budget to hit
🎟️

Loyalty points

  • Held by customers
  • Single redemption cost
  • Expire → breakage / liability
🚚

Fengdou — this work

  • Held by workers; non-tradeable
  • Dual cost: free virtual / budget real
  • Budget target + use-it-or-lose-it ratchet
The new class

A closed-loop, non-tradeable, dual-cost, budget-targeted incentive currency.

Problem SettingA New Problem Class

Problem Setting · 3 / 6

Levers, Timing & Predict-then-Optimize

The firm pulls three levers — but never decides how workers spend:

🎛️

Issuance

how many tokens to mint
→ incentive power & float

🏷️

Real-goods prices

token price of budget items
→ budget burn rate

🎮

Virtual-goods prices

token price of free items
→ the budget valve

The catch: levers are set before workers spend

🎛️Set leversissuance & prices
→commit
🎲Workers redeemuncertain response
→observe
🎯Budget & float realizeddid we hit target?
Predict-then-optimize

The response is unknown at decision time, so the firm must predict how workers will redeem, then optimize the levers against that prediction and the budget. (Bertsimas & Kallus 2020; Elmachtoub & Grigas 2022.)

Problem SettingLevers, Timing & Predict-then-Optimize

Problem Setting · 4 / 6

Dual Cost & the Budget Target with Ratchet

Only the real channel costs money; the virtual channel consumes tokens at zero budget cost:

$$R_t=\sum_{j\in\mathcal R}p^R_j x^R_j+\sum_{k\in\mathcal V}p^V_k x^V_k,\qquad C_t=\sum_{j\in\mathcal R}c_j\,x^R_j\ \ (\text{virtual cost}\approx 0)$$

The objective on annual cost \(C_\tau\) is genuinely two-sided, and underspending is penalized dynamically through a ratchet:

Two-sided penalty

$$\Pi(C_\tau;B_\tau)=\kappa_o(C_\tau-B_\tau)^{+}+\kappa_u(B_\tau-C_\tau)^{+}$$

Overspend infeasible / costly; underspend costly via lost future capacity.

Use-it-or-lose-it ratchet

$$B_{\tau+1}=B_\tau-\rho\,(B_\tau-C_\tau)^{+}$$

Leaving budget unspent permanently shrinks next year's capacity.

Signature

“Hit a target under a two-sided penalty, deciding before demand” = the newsvendor; the ratchet couples years into a dynamic program.

Problem SettingDual Cost & Budget Target

Problem Setting · 5 / 6

Token Float & the Firm's Multi-Objective Problem

Unredeemed tokens accumulate as a float (a contingent liability and an incentive asset at once):

$$K_t=K_{t-1}+I_t-R_t\qquad\text{— non-expiring: it accumulates unless steered}$$

The firm chooses levers, before the shock, to balance three competing goals:

Incentivize

Maximize effort / retention value per yuan of budget.

Hit the budget

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

Stable float

Keep outstanding tokens in a healthy band (Miller–Orr) — avoid hoarding or collapse.

$$\min_{\bm u\in\mathcal U}\ \mathbb E_{\tilde{\bm z}}\big[\,\Pi(C_\tau;B_\tau)-\lambda\,V_\tau(\bm u,\tilde{\bm z})\,\big]\quad\text{s.t.}\quad \underline K\le K_t\le\overline K$$

All of \(C_\tau,\ V_\tau,\ K_t\) depend on the unknown response \(\bm x(\bm u,\tilde{\bm z})\) → solved in two stages (predict, then optimize); the ratchet links years into an MDP.

Problem SettingFloat & Multi-Objective Problem

Problem Setting · 6 / 6

Three Design Principles

The structure of the problem yields three principles that organize the design and that we validate on the data:

P1 · Pressure valve

The virtual price is a near-free valve: lowering it diverts token demand to the zero-cost channel, cutting real-goods cost without destroying token value.

⇒ set issuance to motivate; virtual price to steer the budget.

P2 · Newsvendor target

Levers commit before demand under a two-sided penalty ⇒ aim the budget lever at a critical-fractile target of the cost distribution.

⇒ the valve thins the distribution, easing the target.

P3 · Float & ratchet

Manage the float as a band (Miller–Orr); the use-it-or-lose-it ratchet raises the spending target above its myopic level.

⇒ both embedded in the optimization.

Stance

This is an empirical + optimization paper: P1–P3 are design principles instantiated in the framework and tested on the panel — not stylized theorems.

Problem SettingThree Design Principles

Current Approach · 1 / 4

The Predict-then-Optimize Framework

Predict
issuance forecast
Predict
redemption response
→
Optimize
targets: budget · float
→
Levers
issuance · prices · valve
↻
Feedback
recalibrate

Predict stage

  • Issuance: gradient-boosting forecast of task completions (+ expert estimates for rule-driven tasks)
  • Redemption response: which goods' volumes move with price — inferred without price-experiment data

Optimize stage

  • Set target trajectories for the real-goods budget and the float band
  • Move the three levers to track them; the virtual valve steers the budget (P1)
  • Realized outcomes feed back to recalibrate
Current ApproachPredict-then-Optimize Framework

Current Approach · 2 / 4

Predict: Redemption Response Without Price Variation

Challenge. The internal mall holds prices fixed ⇒ no price variation ⇒ the textbook regression of quantity on price is infeasible. We infer relative elasticity from the composition of demand:

$$Q=\underbrace{N_{\text{buyers}}}_{\text{penetration}}\times\underbrace{q_{\text{per-buyer}}}_{\text{consumption depth}}$$
ClassBuyersPer-buyer qty.Elasticity \(e\)Rationale
AHighHigh(−0.4, −0.1)habit ⇒ inelastic
BHighLow(−1.5, −1.1)substitutable ⇒ elastic
CLowHigh(−0.9, −0.5)sticky ⇒ moderate
DLowLow(−1.5, −1.1)threshold-sensitive ⇒ elastic
Why it suffices

The class drives only the direction of pricing — and we show the policy is robust to elasticity-magnitude error. Same-volume goods can be opposite (A vs. B): composition, not level, sets elasticity.

Current ApproachPredict: Redemption Response

Current Approach · 3 / 4

Optimize: Targets, Levers & the Pressure Valve

Set the targets

A multi-objective score over feasible operating plans selects a target trajectory — budget-channel target, float band, and supporting redemption paths.

$$S^\star=\arg\max_{S\in\mathcal S}\ \textstyle\sum_g w_g\,S_g$$

Move the levers

  • Issuance (reward coeffs): for incentives & float supply
  • Real prices: manage within-channel volume by elasticity class
  • Virtual price = the valve: divert demand to the free channel (P1)
The valve in action

To cut real-goods cost without losing incentive value: lower the virtual price and boost its exposure → token demand flows to the zero-cost channel; real prices fine-tune; issuance is freed to motivate.

Levers are recomputed each period against the targets; this recomputation is what an operator reviews before deployment.

Current ApproachOptimize: Targets, Levers & Valve

Current Approach · 4 / 4

Preliminary Results

Calibrated on an SF Express panel (13 months real records + 5-month validation). Units: \(10^8\) tokens.

IssuanceRedemptionReal-goods (budget)VirtualEnding float
Baseline (no steering)4.443.422.201.223.11
Optimized target4.083.892.261.632.18
Predict-then-optimize4.183.852.201.642.44
Δ vs. baseline−5.8%+12.5%+0.1%+34.8%−21.9%

Budget held, float cut

Real-goods (budget) channel flat (+0.1%) while the float falls 21.9%; float deviation narrows from 43.3% → 12.0%, others within 5%.

Direct evidence for the valve

Virtual redemption +34.8%: the zero-cost channel absorbed the diverted demand — exactly P1. Monte Carlo confirms the outcome is robust to elasticity error (direction invariant).

Current ApproachPreliminary Results

Open Discussion · 1 / 2

Future Direction I — Deeper Empirical Study

Calibrate the response

  • Online price A/B tests to replace literature-inferred elasticities with estimated magnitudes
  • Staged rollouts to learn point-context (vs. cash) sensitivity

Bring in the budget & ratchet

  • Tie the actual ¥ budget and the multi-year ratchet into the numbers
  • Solve the across-year MDP — how the ratchet shifts the interior target

External validity

  • Multi-firm / multi-industry data to test the framework's generality
  • Other incentive currencies: corporate credits, in-game points, platform balances

More benchmarks & tests

  • Counterfactual policy experiments; stress tests under demand shocks
  • Light study of human-operator overrides of the recommended levers
Goal

Move from a validated single-firm calibration to causally identified response estimates and a fully dynamic, budget-and-ratchet-aware policy.

Open DiscussionFuture Direction I — Empirical Study

Open Discussion · 2 / 2

Future Direction II — Robust Satisficing for the Multi-Objective Problem

Limitation now. Targets are selected by a weighted sum of scores — arbitrary weights, no distributional robustness, no risk control over each goal.

Proposal. Recast the multi-objective optimization as a robust satisficing model (Long, Sim & Zhou 2023): give each goal a target and minimize an aggregated risk measure of each goal's shortfall under distributional ambiguity.

$$\min_{\bm u\in\mathcal U}\ \sum_{g}\,w_g\,k_g(\bm u)\qquad\text{s.t.}\quad \sup_{\mathbb P\in\mathcal F\!\left(k_g\right)}\ \mathbb E_{\mathbb P}\big[(\tau_g-V_g(\bm u,\tilde{\bm z}))^{+}\big]\le 0\ \ \ \forall g$$

Goals \(g\in\{\text{budget adherence},\ \text{float band},\ \text{incentive value}\}\); target \(\tau_g\); \(k_g\) = fragility (how fast goal \(g\) degrades as \(\mathbb P\) departs from the empirical \(\hat{\mathbb P}\), with \(\mathcal F(k_g)\) a Wasserstein ball of radius \(\propto k_g\)).

Target-based

Each goal must be met, not traded against the others by hand-tuned weights.

Ambiguity-aware

Robust to the gap between the empirical panel and the true distribution.

Risk-controlled

Minimizing aggregated fragility \(k_g\) keeps every objective robustly satisfied.

Open DiscussionFuture Direction II — Robust Satisficing
1 / 13