The Sealed Envelope

This is the revised and re-rendered version of Episode 85. Two synthetic hosts — Lukas and Estelle, performed with ElevenLabs v3 using the Chris and Cera voice pair — tell a long story about probability, markets, war, forecasting, and the uncomfortable moment when an elegant model meets a physical grid.

The earlier version of this episode had the right ambition but one wrong centre of gravity. It leaned too hard on a too-clean Texas battery payoff: one battery, one perfect forecast, one spectacular hour of revenue. That version was dramatically satisfying and factually too neat. The source-locked version is better because the real story is messier. Forecasts matter. Tails matter. But so do nodal constraints, state of charge, high sustained limits, settlement rules, and whether the asset is physically able to move when the model wants it to move.

That is now the point of the episode.

The argument in one sentence

Almost every important idea in modern markets traces back to a question two seventeenth-century Frenchmen exchanged by mail in 1654 about how to split a pot in an interrupted dice game. Between then and now, the mathematics of probability gets lost, rediscovered, formalized, misused, repaired, and eventually deployed inside physical infrastructure where the future refuses to be a single number.

That is the episode. Everything else is lineage.

The thing that keeps getting lost

The pattern that matters most in this story is not progress. It is neglect.

Gerolamo Cardano writes the first systematic mathematical treatment of probability in Milan around 1564, drawing from a lifetime of gambling. He does not publish it. The manuscript appears almost a century later.

Thomas Bayes writes the first formulation of inverse probability in the 1750s. He does not publish it either. Richard Price finds it after Bayes’ death and sends it to the Royal Society.

Louis Bachelier defends a thesis at the Sorbonne in 1900 that contains much of the mathematical intuition of modern option pricing. It sits largely ignored until the 1950s.

This is one of the strange lessons of mathematical history: foundational ideas do not always arrive with trumpets. Sometimes they arrive in drawers, archives, marginal lives, and papers that the field is not yet able to read.

The sealed envelope

The dramatic centre of the episode is Wolfgang Doeblin.

Doeblin was born in Berlin in 1915, son of Alfred Döblin, the author of Berlin Alexanderplatz. His family was Jewish. They left Germany after the Reichstag fire in 1933 and settled in Paris. Wolfgang studied probability under Maurice Fréchet and became one of the most promising young probabilists in France.

Then the war came.

In February 1940, while serving in the French army, Doeblin sent a sealed manuscript to the Académie des Sciences in Paris. The French tradition of the pli cacheté allowed a scientist to establish priority without making the work public. The title on the envelope was Sur l’équation de Kolmogoroff.

Four months later, on June 21, 1940, cut off from the main French forces in the village of Housseras, Doeblin burned his papers and shot himself in a barn. He was twenty-five.

The envelope remained closed for sixty years.

When it was opened in May 2000, it contained major work on diffusion processes, including a version of the change-of-variables formula later associated with Kiyosi Itô. It was not a clean replacement of the Itô story. It was something more moving: overlapping insight, developed under impossible conditions, left unread because history intervened.

Ito, the extra term, and finance

On the other side of the world, Kiyosi Itô was working in wartime Japan with limited access to European mathematical literature. His landmark 1944 paper on the stochastic integral becomes one of the pillars of twentieth-century mathematics.

The key idea is simple to say and difficult to build: ordinary calculus breaks on Brownian motion. A smooth curve becomes locally straight when you zoom in. Brownian motion does not. It remains jagged at every scale.

That is why Itô’s formula has an extra term. In ordinary calculus, the change in a function depends on the first derivative. In stochastic calculus, the roughness of the path contributes a second-order term that does not vanish. That extra term becomes the quiet engine under modern diffusion finance.

Black, Scholes, and Merton use this machinery in 1973. Their formula is famous, but the more important idea is not the formula. It is the hedging argument. Option pricing is not primarily a prediction about where the stock will go. It is a no-arbitrage replication argument.

That leads to the conceptual centre of the episode: P and Q.

Forecasting happens under the real-world probability measure, P. Pricing happens under the risk-neutral measure, Q. The market can be terrible at predicting returns and still price derivatives coherently. A volatility implied by an option price is not simply “what traders think will happen”. It is the volatility that makes a no-arbitrage pricing equation fit the market.

Almost every confused piece of popular writing about derivatives misses that distinction.

Tails, overfitting, and model risk

The episode then follows the machinery into the twentieth and twenty-first centuries: martingales, Black-Scholes, GARCH, volatility smiles, Mandelbrot, rough volatility, backtest overfitting, limit order books, neural SDEs, and AI systems in finance.

The recurring warning is the same in different forms:

Markets are wilder than convenient models want them to be.

Mandelbrot saw it in cotton prices in 1963. Long-Term Capital Management demonstrated it in 1998. The 2008 crisis demonstrated it again at global scale. Modern quantitative research faces the same problem in a different costume: if you try enough backtests, one of them will look brilliant by accident.

The discipline is not “use the fanciest model”. The discipline is honest evaluation: out-of-sample testing, calibration, cost-aware backtesting, stress testing, and admitting that a model can be impressive in research and fragile in production.

That is the bridge to energy.

The Texas morning

The closing act now centres on February 19, 2025 in ERCOT.

There was a large net-load forecast miss: demand minus expected wind and solar came in more than 17 GW away from the day-ahead view. There was local congestion around Round Rock. At the RHESS2_ESS1 settlement point, associated with the Rabbit Hill battery near Georgetown, locational marginal prices averaged $28,187/MWh between 8 and 9 a.m.

The simple version would be: a battery saw the tail, held energy, discharged at the top, and made a fortune.

The real version is more useful.

GridStatus later looked at the 60-day ERCOT data and found that Rabbit Hill did not simply capture the highest-price intervals. During much of the worst congestion, its high sustained limit was zero. In plain English: the asset was not available to sell into the most extreme part of the event.

That is the better lesson. Probability is necessary, but it is not sufficient. A forecast distribution does not dispatch a battery by itself. The grid is not a spreadsheet. It has topology, telemetry, constraints, market rules, asset state, and physical limits.

The correct payoff is not “one model predicted the spike”.

The correct payoff is: do not trust point forecasts; do not ignore tails; do not confuse a model with the physical system it is meant to control.

Why this is for the team

I wanted this episode on the record because the real subject is Sourceful’s operating problem.

Every battery we coordinate is part of a probabilistic decision system. The inputs are uncertain: weather, price, congestion, availability, degradation, human behaviour. The outputs have to be both economically intelligent and physically executable: bid, hold, discharge, reserve, stop, wait.

That is why the old probability lineage matters. Pascal and Fermat, Bernoulli, Bayes, Bachelier, Kolmogorov, Doeblin, Itô, Black-Scholes, Mandelbrot, Gneiting — none of them were writing software for distributed energy assets. But the mathematics they built is part of the language we need to operate those assets honestly.

We are doing financial engineering on a physical system. The financial engineering only works because the math was built. The physical system only works if the math is connected to telemetry, constraints, asset state, and control.

Key Takeaways

* Probability begins, in its modern form, with gambling problems and correspondence, not with physics or finance.

* A surprising amount of the history of probability is a history of being ignored: Cardano, Bayes, Bachelier, and Doeblin all wrote work that reached the field late.

* Doeblin’s sealed envelope is one of the most moving documents in twentieth-century mathematics: work on diffusion processes written in wartime, deposited in Paris, and opened only in 2000.

* The P/Q distinction is the conceptual keystone of modern finance. Forecasting and pricing are not the same problem.

* Mandelbrot’s warning about tails remains alive. Models that assume the world is smoother than it is will fail exactly when they are most needed.

* Modern AI forecasting is not a replacement for probabilistic discipline. It makes calibration, uncertainty, and evaluation more important, not less.

* The ERCOT February 19, 2025 event is useful precisely because it is messy. Forecasting, constraints, location, state of charge, HSL, and settlement all matter.

* For Sourceful, the lesson is concrete: coordination is not just prediction. It is prediction connected to a controllable physical asset in time.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com