W
WhiteBeard/Direction
Leverage-sizing research
Direction

Where this is going

The goal, the method we agreed, and an honest read on how far along each part is.

The goal

One number: X — how much a hedge fund can safely lend against a broker's first-loss capital.

“Based on this broker's actual order flow, at X = 4 your capital would not have been touched in 20 years, with 99.9% confidence.”

That sentence, defensible to a fund's risk desk, is the deliverable. Everything else is machinery to produce it honestly.

The two layers

The work splits cleanly. One half is settled; the other is the real research.

Decision layerDone

Turns a worst-case loss into a lendable multiple. It reduces to one relationship, and it reproduces the Forex Hub answer (X = 4) from the call.

X = 1 / (λ × g) − 1
Estimation layerIn progress

Produces g — the worst loss the book can actually take — from the broker's own flow. This is the hard part, and it is the six stages below.

5 of 6 stages complete · 6 running

The method

The engine for estimating the worst-case loss, as laid out on the call.

  1. 01
    Reconstruct the inventory

    Rebuild the position held at the LP, per currency, from the Pawn's A-book flow. Verified: it reproduces the platform's own risk number within 5% on 105 of 136 days.

    Done
  2. 02
    Replay it across history

    Hold each of 55,627 minute-level positions fixed and drag it through 26 years of history. Worst one-day loss: 8.5–10.5% of the position, on both books. Every binding day is a 2026 gold day.

    Done
  3. 03
    Correlation & randomness

    Rolling 11×11 correlation over 26 years (it is a time series — XAU/AUD ranges −0.33 to 0.79) plus 20,000 correlated random scenarios per snapshot. At today's volatility the 99.9% loss is 4.3–5.7%; the unconditional replay stays the buffer.

    Done
  4. 04
    Fit the drawdown formula

    Three forms fitted on 70% of history, tested on 30%. The scenario formula — max over the 36 historical days that ever bind — is exact out of sample. The linear reading: 7.8% per $ long gold, 6.1% per $ short.

    Done
  5. 05
    Max-of-max → first-loss

    C₁ = $3.27M on the retained book ($192M if the hedged book is charged — see the open question). X follows: 4.9× at 2:1, infeasible above ~10:1.

    Done
  6. 06
    Instantaneous shock check

    Top-20 sub-30-minute shocks per asset since 2015, mined from 5-minute bars (the 2015 gold flash crash, Brexit, the 2016 election), applied jointly to every snapshot. The curated PiFolio list will replace the mined one when it arrives.

    In progress

What we need to keep moving

  • The hedge-offset answer.Charging the hedged (A-book) position gives a first-loss of $192M; charging the retained (B-book) risk gives $3.3M. The A-book is a hedge — its loss is the client book's gain — so we size on the B-book and report the A-book notional as the financed exposure. Does a margin call on the hedged leg consume the broker's capital before clients settle? The multiple X is the same either way; only the dollar buffer moves.
  • The PiFolio event list — the curated instantaneous shocks. Stage 6 runs today on shocks mined from 11 years of 5-minute bars; the curated list is a cross-check, not a blocker.
Next: fund-facing report template