Did our hedge do what we intended?

Answering that question means pulling four systems together by hand, in a spreadsheet, under a deadline. Every single time.

See what a teardown returns

Variable annuity & indexed guarantee hedging

01The residual

Your hedge underperformed by $1.5M on Tuesday.

Liability moved +$8.0M. The hedge returned +$6.5M. Decomposing both sides and setting them against each other is the only way to learn where the difference came from — and how much of it nobody can account for.

Factor	Liability	Hedge	Gap	Reading
Equitydelta	+5.0	+4.7	−0.3	Basis risk — policyholder funds fell further than the proxy index
Ratesrho	+2.0	+1.9	−0.1	Curve bucket mismatch
Volatilityvega	+1.2	0.0	−1.2	Unhedged. Deliberate, or missed? The number alone cannot tell you
Timetheta	−0.8	0.0	+0.8	Expected. Funded by fees
Behaviorlapse / withdrawal	+0.1	0.0	−0.1	Not hedgeable
Unexplained	+0.5	−0.1	−0.6	No one can say where this came from
Net	+8.0	+6.5	−1.5

USD millions · illustrative

The vega gap is the largest line — but it may be a decision, not a failure. You cannot tell from the number. You have to read the hedging policy.

The residual is the second largest line, and it is not a decision. It is the part your model did not predict. That gap is what this product exists for.

02The seams

The calculations already exist. The connections don’t.

No insurer is short of calculation engines. The problem is that each one is authoritative inside its own boundary and blind immediately outside it. The answer lives in the space between them, so a person reassembles it by hand.

Actuarial

Liability greeks, guarantee value, cohort structure

Cannot see what the desk actually traded

Asset & Treasury

Hedge positions, realized and mark-to-market P&L

Cannot see the liability it is hedging

Execution

Intended specification versus filled orders

Cannot see the risk impact of the difference

Policy Admin

In-force movement, new business, surrenders

Cannot see any of the above

We do not build another engine. We live in the seams.

03Why it’s capital

Unexplained is not an annoyance. It is capital.

Under VM-21, reserve credit for a hedging programme depends on an error factor E, which ranges from 5% to 100%. A lower factor means a smaller reserve.

To justify it, you must demonstrate — with at least twelve months of experience and back-testing — that the model can replicate actual hedging results. Effectiveness is measured, documented and reported to the Risk Committee and Appointed Actuary no less than quarterly.

A model that cannot explain the past is not a documentation problem.

It is a reserve you did not have to hold.

04How it works

Explain yesterday first. Then earn the right to warn about tomorrow.

Prediction cannot be checked on the day it is made. A teardown can — you already know what happened. So the product starts by taking the past apart, and turns what it finds into what gets watched.

1
Portfolio comprehension
We read your files and build a model of your exposure structure — which products sit on which funds, mapped to which market factors, hedged with which instruments. Column meanings, sign conventions, units and as-of rules are inferred, proposed, and confirmed by you. Then frozen into deterministic code.
files → exposure model
2
Teardown analysis
Liability movement and hedge P&L are decomposed by factor and set side by side. What remains is the residual — reported as a daily series, not a single number. Every figure traces back to a source file, field and formula.
period data → decomposition + unexplained
3
Signal definition
Recurring loss paths found in the teardown become watch rules. You refine them in conversation with an agent that argues from your own evidence, and proposes what to add, patch or retire. Thresholds come from your residual distribution, not from us.
findings + dialogue → signal set
4
Continuous ingestion
Market data we collect ourselves. Your internal extracts arrive on a schedule. Receipt time and as-of time are recorded separately — because batch lag is itself a source of residual.
feeds → time series
5
Evaluation & flagging
Incoming data is checked against the signal set. Breaches raise a flag ranked by distance to a decision boundary, not by size of move. The next teardown scores the last quarter's signals, and the thresholds adjust.
data × rules → flags

The loop closes on itself. The next teardown grades the signals the last one produced — precision and recall become measurements, not claims, and the thresholds move accordingly.

05Boundaries

What we don’t do.

We don't calculate your greeks

Numerical truth comes from your validated engines. We consume their output; we never replace it.

We don't recommend trades

We tell you what to investigate and who has to decide. What to execute is yours, and stays yours.

No language model touches a number

Decomposition, regression and residuals run in deterministic code. Same inputs, same outputs, always auditable.

Models infer how your files are shaped, read your policy documents against the numbers, and argue with you about which signals to keep. They sit at the entrance and the exit — never in the middle.

06Next step

Send us last quarter. We’ll take it apart.

No integration project. Four extracts you already produce. You know what actually happened last quarter — so check our decomposition against it, line by line.

Request a teardown

What we need

Liability greeks
actuarial extract, daily
Hedge positions & P&L
asset / treasury, daily
Fund-to-index mapping
current and prior estimation
Hedging policy / CDHS
document, optional but useful

Market data is on us. CSV or Excel is fine.