We consider a futures contract written on a non-tradable KPI (e.g., quarterly vehicle deliveries), with discrete realization and a bounded payoff $L(K) \in [0, L_{\max}]$.
Unlike traditional futures, tail risk is bounded and price discovery is sparse — KPI contracts typically have as few as 2 historical observations, rarely more than 20. Margin must therefore address:
Let observed KPI levels be $K_{t-n}, \dots, K_t$ for however many periods of history exist (minimum 2; here 8 quarters). Define relative quarter-over-quarter changes:
Relative returns are used (rather than levels) for scale invariance and improved stationarity. For TSLA-Q2-2026-DELIVERIES, the 7 observed returns span Q2 2024 – Q1 2026, capturing the full recent delivery cycle including the Q1 2025 miss (−40%) and Q3 2025 recovery (+37%).
| Quarter | Deliveries | Return $r_i$ |
|---|---|---|
| Q2 2024 | 443,956 | — |
| Q3 2024 | 462,890 | +4.3% |
| Q4 2024 | 495,570 | +7.1% |
| Q1 2025 | 296,751 | −40.1% |
| Q2 2025 | 363,000 | +22.3% |
| Q3 2025 | 497,120 | +36.9% |
| Q4 2025 | 418,227 | −15.9% |
| Q1 2026 | 336,681 | −19.5% |
With $n \ge 2$ observations (use however many quarters of history exist; here $n = 7$), compute sample mean and unbiased sample variance:
For the TSLA contract: $\hat{\mu} \approx +2.0\%$ per quarter, $\hat{\sigma} \approx 26.5\%$ per quarter.
With small $n$, $\hat{\sigma}^2$ is a downward-biased estimator of true tail risk: extreme events are under-sampled, estimation error is large, and naive ES underestimates actual risk. We apply a multiplicative inflation:
This can be viewed as Bayesian shrinkage toward higher variance, or as a finite-sample correction for parameter uncertainty. As $n \to \infty$, $\sigma_{\text{adj}}^2 \to \hat{\sigma}^2$.
We model the next-period KPI return as a scaled and shifted Student-$t$ distribution:
The Student-$t$ with $\nu \in [4, 6]$ degrees of freedom provides heavier tails than Gaussian, capturing the jump risk characteristic of KPI outcomes (earnings misses, supply disruptions, macro shocks). The Gaussian assumption would materially underestimate tail losses for this contract.
| Parameter | Meaning | Value |
|---|---|---|
| $\nu$ | Degrees of freedom (tail thickness) | 5 |
| $\hat{\mu}$ | Mean quarterly return | +2.0% |
| $\sigma_{\text{adj}}$ | Inflated quarterly volatility | 33.2% |
The settlement-period KPI level is projected from the most recent observed level $K_t$ (Q1 2026: 336,681):
The projected settlement is clamped to the contract's bounded range $[L_{\min}, L_{\max}]$ for payoff calculation. Note that $K_{t+1}$ is the simulated underlying, not the contract price — the market price (current consensus: 460,000) reflects the market's expectation of this quantity.
For a bounded linear (continuous-range) KPI futures contract, the dollar loss on a position is:
where $\delta = +1$ for BUY, $\delta = -1$ for SELL, $P$ is the trade price, $N$ is the dollar notional, $R = L_{\max} - L_{\min} = 300{,}000$ is the range width, and $\tilde{K} = \text{clamp}(K, L_{\min}, L_{\max})$. The key property is:
We use Expected Shortfall (ES, also called CVaR) rather than VaR as the primary risk measure. First, define Value-at-Risk:
Then Expected Shortfall at confidence $\alpha$:
ES is preferred over VaR because it is a coherent risk measure (sub-additive, rewarding diversification), sensitive to tail severity, and more robust to distributional misspecification. It is the standard for margin under CFTC/EMIR clearing rules.
In implementation, ES is computed via Monte Carlo: 12,000 samples drawn from $t_5(\hat{\mu}, \sigma_{\text{adj}})$ with a fixed seed for determinism.
With few data points (possibly as few as 2), the ES estimate is statistically unstable and tail behavior is under-identified. A stress overlay supplements the statistical measure with deterministic scenarios:
| Scenario | Return | $K_{\text{sim}}$ | Comment |
|---|---|---|---|
| Historical min | −40.1% | ≤ 300,000 | Q1 2025 actual |
| Historical max | +36.9% | 460,000+ | Q3 2025 actual |
| $-\lambda\sigma_{\text{adj}}$ | −116% | ≤ 300,000 | hypothetical extreme |
| $+\lambda\sigma_{\text{adj}}$ | +116% | ≥ 600,000 | hypothetical extreme |
The final risk measure uses $\max(\text{ES}_\alpha, \text{StressLoss})$ so either constraint can be binding.
KPI futures have a fundamentally different time structure from standard futures. The correct requirement is:
Margin must increase as settlement approaches and converge to the maximum possible loss at the event date. This is the opposite of diffusion-based VaR scaling ($\sqrt{\tau}$), which collapses toward zero at maturity.
We construct a deterministic convergence term that starts at zero when the contract is listed and grows to $L_{\max}$ at maturity:
Properties:
The parameter $k > 0$ controls the steepness of convergence. Large $k$ produces faster early growth; small $k$ produces a more gradual ramp.
For the TSLA contract at 68 days remaining ($\tau = 68/182 \approx 0.374$): $\text{convergenceTerm} = L_{\max} \cdot (1 - e^{-1.5 \times 0.626}) \approx 0.607 \cdot L_{\max}$.
Large positions relative to market depth impose liquidation cost and adverse-selection risk not captured by the statistical ES. An add-on penalizes concentration:
| Parameter | Meaning | Value |
|---|---|---|
| $\gamma$ | Concentration coefficient | 0.05 |
| $Q$ | Position dollar notional | trade-specific |
| $D$ | Market depth proxy | $500,000 |
For a $10,000 notional position, ConcAdd $= 0.05 \times (10{,}000 / 500{,}000) \times L_{\max} \approx 0.1\%$ of $L_{\max}$ — negligible at normal sizes, material above $D$.
subject to $\text{IM}(\tau) \le L_{\max}$. The formula has three components:
At listing ($\tau=1$): $\text{IM}(1) = 0 + \beta \cdot L_{\max} + \text{ConcAdd} \approx 30\% \cdot L_{\max}$ — margin equals the statistical floor. At maturity ($\tau \to 0$): the convergence term dominates and the cap binds, so $\text{IM}(0) = L_{\max}$.
| Property | Classical futures (VaR) | KPI futures (this model) |
|---|---|---|
| Risk type | diffusion | event-resolution |
| Time scaling | $\sqrt{\tau}$ → 0 at maturity | $(1-e^{-k(1-\tau)})$ → $L_{\max}$ |
| IM at maturity | 0 (liquidated before) | $L_{\max}$ (outcome imminent) |
| Framework | liquidation horizon | information horizon |
The final initial margin has a single hard cap:
The statistical term is floored at $\beta \cdot L_{\max}$ (30%), which acts as the minimum initial margin at listing. At maturity the convergence term alone is sufficient to drive total margin past $L_{\max}$, so the cap always binds at $\tau = 0$, ensuring $\text{IM}(0) = L_{\max}$.
Production calibration for TSLA-Q2-2026-DELIVERIES:
Parameters should be reviewed at each contract listing. Contracts on higher-frequency or more liquid KPIs (e.g., monthly web traffic) warrant smaller $c$ and $\lambda$; contracts on CEO-sensitive or policy-driven KPIs warrant larger values.
ES validation: verify that realized conditional tail losses match the model:
Exception rate: the breach probability should be stable and close to $1 - \alpha$:
Stress coverage: stress scenarios should bound realized extreme outcomes:
Given the small settlement history, backtesting must rely on cross-sectional data from similar KPI contracts (other automakers, comparable cyclical industrials) until sufficient TSLA settlement history accumulates.
Mitigations: stress overlay provides non-parametric protection; variance inflation with $c=4$ conservatively shifts the distribution; $\beta$-floor prevents collapse near maturity.
The bounded payoff structure fundamentally changes the margin problem. Because $L \le L_{\max} < \infty$:
This removes unbounded tail risk and default cascade dynamics. Clearinghouses on traditional futures must set margin at a VaR percentile and accept residual tail exposure; here the worst-case loss is analytically computable. The result is a framework that is:
Bayesian volatility model: place a hierarchical prior over $\sigma^2$:
The posterior predictive is a Student-$t$ distribution with degrees of freedom $2\alpha_0$ — a clean link to the parametric framework above, with $c$ emerging naturally from the prior parameters.
Cross-sectional volatility pooling: share variance information across related contracts (e.g., RIVN, XPEV deliveries):
Regime-switching: model normal vs. stressed quarters explicitly:
Regime detection could use macro signals (supply chain PMIs, consumer sentiment) to shift the mixture weight dynamically.
Intra-quarter updates: incorporate monthly delivery flash estimates as soft information to update $K_t$ prior to formal settlement, allowing margin to tighten in real time as the outcome resolves.