The lender-facing note: how to size debt on these forecasts, and exactly how to treat the published tail failure.
Underwriting with an honest model
A note for lenders, credit committees and their technical advisers: what a
CECadence forecast is for in a credit process, the sizing procedure we
recommend, and — before your diligence finds it — exactly what our published
out-of-sample record says, including the failure. Every model number cited
here already appears on a published surface: the
held-out note, the
pre-registration record and the
quarterly index report. Nothing is recomputed for
this page.
1. What the model is for in a credit process
CECadence forecasts GB battery revenue at the cannibalisation
equilibrium: wholesale prices and the whole fleet's dispatch are solved
jointly, so the revenue line already reflects the future fleet competing away
the spreads it earns from. For a lender this is the relevant number — the
fleet a 2027+ project operates against is the built-out fleet, not today's.
Every run also reports the naive price-taker figure: the number an
exogenous-price forecast produces, and the industry default. It is shown for
comparability with other vendors' outputs, and it is systematically
optimistic — at 2030 fleet sizes a price-taker forecast overstates revenue by
100–300%+. The gap between the two figures is reported on every run as the
cannibalisation bias.
For candidate financings the model produces a multi-year revenue track
against the growing fleet, with the minimum DSCR and its binding year
identified. The credit question is a path question, not a snapshot.
2. The procedure
One convention first, because it prevents sign errors: revenue bands use the
exceedance convention — PXX is the value exceeded in XX% of Monte Carlo
draws, so P90 is the conservative (low) band.
Size the debt on the P50 equilibrium revenue track, at the published
1.40× P50 DSCR target. The P50 of the cannibalisation-aware track is
the planning number.
Test covenant headroom on the P90 (exceedance) band, at the
1.10× P90 DSCR covenant. The P90 band is conservative by construction
and does not depend on the model's upside tail.
Underwrite the minimum-DSCR binding year of the multi-year track, not
a single-year snapshot. Cannibalisation makes the revenue path
front-loaded relative to a flat extrapolation; the binding year is where
the credit lives.
Never underwrite on the naive price-taker figure. It is published so
you can see how far the industry-comparable number sits above the
equilibrium one — that distance is the risk, not the opportunity.
Give no credit to modelled upside tails. This is not generic prudence;
it is what our own published record instructs, as the next section sets
out.
3. What our validation record says, straight
CECadence validates on pre-registered held-out windows: the rules are
frozen in dated commits before the test data is fetched, and the grades are
published whatever they say. Two held-out windows have been graded so far.
The published thresholds are mean ≤ 5%, P50 ≤ 10%, P95 ≤ 20% (price error vs
realised GB prices).
Held-out window
Basis
Mean
P50
P95
Published verdict
2025 (full year)
synthetic
+12.0%
8.2%
59.4%
FAIL — mean and scarcity tail
2025 (full year)
realised
1.6%
24.0%
55.4%
best mean of any year; FAIL on P50/P95
2026H1 (window)
synthetic
1.4%
23.0%
55.6%
window-mean PASS; FAIL on P50/P95
Read as a credit analyst would:
The annual mean — the revenue level — is the model's strong claim. The
held-out 2026H1 window mean came in at 1.4%, the first out-of-sample
mean pass in the record, and the realised-basis 2025 mean (1.6%) is the
best of any year. We publish the fragility alongside the pass: the 2026H1
mean's 90% bootstrap confidence interval (0.2–6.0%) spans the 5% target, so
we call it a pass on the point value and a fragile one — one
favourable-regime window, not a general accuracy claim.
The P95 scarcity tail runs hot out of sample, and we publish that
failure. In both held-out windows the modelled extreme-price tail
overstated the realised one. The practical meaning is the instruction in
step 5 above: do not lean on modelled upside tails. The sizing procedure is
deliberately robust to exactly this known weakness — debt is sized on the
P50 level and covenanted on the P90 low band, and neither leans on the
upside tail the record grades as failing. Where the central case has missed
out of sample it has missed cold: on the 2026H1 window the model's P50
sat below the realised value (£74.0/MWh modelled vs £96.1 realised) — an
under-statement, which for a lender is the conservative direction. The
over-statement stays confined to the upside tail a prudent financing never
relies on.
The tail failure is a disclosed data limit, not a flaw tuned away. The
public record contains one genuine scarcity year (2022). A tail curve
cannot be fitted to be right for both crisis and calm regimes from a single
crisis observation; four re-shape attempts were pre-registered, fitted, and
rejected under their own frozen adoption rules rather than adopted to make
the table look better. The fix is more crisis-regime data, and until it
exists the weakness stays published, not patched.
4. Why you can audit this
Pre-registration with dated commits. Every validation rule is frozen in
a committed protocol before the data it is judged against is fetched, and
the grades publish pass or fail — the record above includes failures and
rejected model changes. The public record, with freeze commits and dates:
/heldout-protocol.html.
A measured benchmark from open primary data. The CE GB BESS Index is
computed from Elexon settlement and NESO auction records under their open
licences — no third-party benchmark licence — with every convention and
every exclusion quantified, so an analyst can verify it line by line:
/quarterly/latest.html.
A standing out-of-sample record from 2026Q2. Each newly settled index
quarter grades the frozen revenue parameterisation before any recalibration
may use that quarter's data. The record accrues quarter by quarter and,
once started, is never restarted.
Dossiers for signed-in diligence. The validation dossier
(/validation.html) and the limitations register
(/limitations.html) are available to signed-in users —
a free account suffices. The full pre-registration protocol, the document
of record, is available on request.
The diligence pack
For the walk-through — the protocol commits, the evidence files, and a
line-by-line reconciliation of any number on this page — email
hello@compoundingenergy.com.
Forecasts are projections, not advice. CECadence outputs are model
forecasts for use inside your own credit process; they are not a
recommendation to lend, invest or transact.