Sensitivity analysis

A NAV is a point inside a space of uncertainty. Every valuation therefore ships with a tornado analysis: the engine re-runs with one lever perturbed at a time, and the resulting deltas are published next to the base case. The tornado freezes into every published version, each row recording the exact perturbation it applied, so a valuation permanently declares the uncertainty known when it was issued.

The analysis answers three questions: which input moves the NAV most per unit of change (where model risk lives), what the NAV would be under stated alternative conditions, and what severe-but-plausible cases imply for protocol buffers and mechanism design.

Directionality

Each lever belongs to one of three types, which determines its justification and whether it runs in one direction or two.

Quantiles test both tails of an observable distribution, at empirically calibrated magnitudes that may come out asymmetric. Parameters test both ends of a defensible disagreement range, symmetric by construction. Scenarios test a discrete event in its single real direction, since a regime change has no meaningful mirror image.

Calibration of the quantile levers

Production. Monthly specific production across the portfolio (1,394 project-months, 30 projects, validated invoices) shows a pooled relative dispersion of σ_month ≈ 20%. Annualizing under bounds for the serial correlation of weather, σ/√12 if independent and σ/√6 under moderate autocorrelation, gives σ_annual ≈ 5.7 to 8.1%; the model uses 6%. The P90/P10 pair follows as ∓ 1.28σ ≈ ∓ 8%, recalibrated annually as telemetry accumulates.

Exchange rate. FX quantiles are calibrated per portfolio currency against its official exchange-rate series. For the current portfolio (COP), rolling 12-month changes over the full available history (1991–2026, 405 windows) have quantiles Q_10 = -12.1%, Q_50 = +4.8%, Q_90 = +24.4%. The model's PPP path already carries the secular depreciation drift, which sits near the median, so the sensitivity shock is the quantile centered on the median: devaluation +20%, appreciation −17%. The asymmetry is in the data, not assumed: devaluations jump, with crisis episodes reaching +37% in a year, while appreciations crawl.

The lever set

Lever	Type	Magnitude	Basis
Production P90 / P10	quantile	−8% / +8%	Own-portfolio telemetry, 1.28σ
FX devaluation P90 / appreciation P10	quantile	+20% / −17%	Official exchange-rate series of the portfolio currency (COP: 1991–2026), median-centered
Discount rate	parameter	±100 bps	Width of the defensible range on the rate build-up
OPEX	parameter	±20%	Interim range, to be replaced by backtesting statistics
Import tax on replacements	scenario	full local VAT + duty (19% + 5% in the current portfolio)	Loss of the renewable-energy import exemption
Certificate price	scenario	→ 0	Collapse of the voluntary certificate market
Combined adverse / favorable	scenario	production ∓8%, rate ±100 bps, OPEX ±20%	The three continuous levers stressed jointly, both directions

Combined scenarios and the NAV range

The combined rows are full engine runs with all three perturbations applied at once, not sums of individual deltas. Interactions matter: a leaner cash flow bears the inflated OPEX proportionally harder, while a higher rate discounts flows that are already smaller. For a representative project, the individual adverse deltas of −11.2%, −7.4% and −7.8% sum to −26.3%, but the joint run yields −24.8%. Only the full run can say which way the interactions cut.

The pair brackets the NAV in a declared range; the same representative project, with a base NAV of 1.096, brackets to [0.824, 1.416]. The adverse bound informs buffer sizing, the favorable bound completes the picture for both sides of issuance and redemption. Discrete events stay out of both combinations, since stacking regime changes onto a bad year stops being a stress test, and FX is reported alone because its correlation with the other levers is ambiguous: devaluation episodes in the portfolio's markets tend to coincide with local inflation that the PPA re-indexes.

Symmetric inputs still produce asymmetric outputs. Present value is convex in the discount rate, so −100 bps gains more (+8.5% in the representative case) than +100 bps loses (−7.4%). Publishing both sides makes the convexity visible.

What this analysis is not

The tornado is not a confidence interval on the NAV. A joint interval would need a correlation structure across all inputs that no available data supports. Per-lever quantiles, calibrated on observable data and individually checkable, say more than a simulated distribution built on correlations nobody can verify.