Their own volume kept blindsiding them.

On an options platform, monthly volume is not a slide. It is servers, headcount, liquidity buffers, and whether risk walks into the month calm or already behind. Their existing models were the usual kind: simple, cheap, and wrong in the ways that hurt. High months arrived as surprises. Quiet months still got staffed like busy ones.

A neural net was on the table. It did not survive the conversation. Nobody on that floor was going to tell a regulator, or their own operations lead, “the network said so.” They needed a forecast they could take apart. Which inputs. How sure. What happens if the market changes.

What we ran

Years of their own history, plus outside prices: futures, volatility, a few macroeconomic indicators, the calendar. Train on the earlier stretch. Score on a later stretch the model never saw. If it only worked on the data it trained on, it was not going in.

On that later stretch it was 41% less wrong than what they had. During the test window, every major volume spike showed up in the forecast before it showed up in the actuals.

The thing they already suspected

Futures were doing most of the work. More than 70% of the signal sat in that one family of prices. Volatility, macro, calendar: real, smaller. The team had felt this for years. They had never seen it measured cleanly enough to bet the staffing plan on it.

That is the part people skip. Accuracy gets you a meeting. A driver you already half-knew, quantified, is what gets the model into production. You can say “volume follows the futures complex” in a standup without sounding like a magician.

They did not stop at volume

The first model went live. Then they asked for five more: active versus idle accounts, liquidity they would need, operational metrics, additional volume, and margin. Same method. Forecasts landing in the internal tools and Power BI they already lived in, not a new system to learn.

That is the review that matters. Not a quote. A reorder.