In March I published a piece called “The Holy Grail of Volatility Modelling: What Happens When You Actually Try to Build It.”

The title is the editorial position. Not what the literature says a joint SPX/VIX calibration ought to do. What happened when I built one and ran it.

Mostly it failed, across two parts and ten findings. Heston pinned against its correlation boundary at rho = -0.99 during live joint calibration. The Bates stochastic volatility jump model degenerated on daily data. A hidden Markov regime classifier handled the crises it had seen and fell over on the ones it had not. Some of it worked: path-dependent volatility beat GARCH by a factor of two, and an XGBoost regime classifier reached 86.95% accuracy. I published all ten in order, including the ones that ended in a wall.

That is the publication, and it is worth being precise about why.

There is no shortage of people publishing edge claims. There is a severe shortage of people publishing the part where someone tried to break one. So the work here is not finding the anomaly. It is establishing whether the anomaly survives contact with transaction costs, parameter instability, and the recalibration schedule a real desk actually runs. Often it does not, and that is the finding.

Two recent examples.

In July I wrote up the Hull-White cadence gap, where volatility is recalibrated daily and mean reversion roughly monthly. The gap is real and independently documented. The last section of that piece establishes that exactly one vendor, one currency and one six month window has ever shown it producing a measurable valuation error, and that nobody has shown that error converts into money. Including me. I could have stopped one section earlier and had a much better headline.

And “I Added Transaction Costs to My Goldman Sachs Backtest. The 49% Return Became a -9.6% Loss.” I broke my own backtest in public, in detail, because the costs were not a rounding adjustment. They were the entire result.

Nobody pays for that part at the point of production. It is the part almost everyone skips, and skipping it is why so much published work does not survive contact with a real book. It is why a piece here takes days instead of hours, and it is the honest reason for a paywall.

Paid subscriptions are now live

I turned them on 27 July and said nothing. Twelve people found them anyway. Thank you. You paid before I asked, which is a stronger signal than anything I could have written here.

What stays free

My most recent three months of posts. Always. It is a rolling window rather than a one-time unlock, and it currently holds 38 pieces going back to late May. Anything older moves behind the paywall automatically as it ages out.

So a reader arriving cold gets three months of current work and can judge it on the work. That does not change, and there is no version of this where it does.

What paid gets

The archive. 251 posts since September 2025, 213 of them now behind the wall. The volatility modelling series in full. The Heston and three-dimensional vol surface work. The Avellaneda-Stoikov inventory skew analysis, the OLS bias in mean reversion speed estimation, the Squarepoint and Man AHL teardowns.

Continuing deep dives on volatility surfaces, model calibration and systematic strategy design. Same standard: sourced, reproducible, and willing to state plainly when a result does not hold.

Comments and questions on every post. Discussion is closed at the moment and I answer by email, inconsistently and usually late. Open threads put the answer where it is worth something to more than one person. If you want to tell me a calibration finding is wrong, I would rather you did it where everyone can read the argument.

The cadence

Two to three in-depth pieces a week. Treat that as a commitment rather than an aspiration, and note that it is not a stretch: it is what the last six months already measure at, just under three a week. Every post is dated and public, so if I miss it, you will be able to see that I missed it.

The price

$10 a month. $100 a year, which works out to two months free. $300 for a founding membership, if the work has already been worth more to you than list price and you want to say so.

One thing plainly: none of this is investment advice. There are no signals, no recommendations, no positions, and no track record on offer. It is research and education about how models, desks and strategies work, including the many places where they do not.

If you stay on the free list, nothing changes for you. You keep getting every new piece for three months, the same as you have been. No campaign, no countdown, no second announcement.

To the twelve who paid first: this became a real thing partly because you treated it as one before I did.

Cover photograph: Kidfly182, CC BY-SA 4.0, via Wikimedia Commons.