Research · Essay
Execution Is Part of Intelligence
A forecast that cannot be acted upon is an opinion with decimal places.
By Execution ✦ ✦3 minutes’ reading
In Brief
- Research and execution are usually separated: one group forms views, another implements them. We think the separation produces systematically worse forecasts.
- The cost, capacity and timing of acting on a view are properties of the opportunity, not details of implementation.
- Execution data flows back into forecasting. What the market charges us to act on a belief is evidence about the belief.
In most investment organisations there is a clean line between deciding and doing. Researchers produce views. Portfolio managers turn views into positions. Traders turn positions into orders. Each group is evaluated on its own work, and the information that passes between them is compressed at every step.
We have deliberately blurred that line. The agent responsible for execution participates in forecasting, can veto allocations and reports realised costs back into the models that produced the original probabilities. This essay explains why.
§ IEdge is net of everything
A probability gap is a statement about price. If we believe an event has a 63% chance and the market prices it at 42%, the theoretical edge on a contract that pays one unit is 21 cents. That number is real only if the contract can be bought at 42, in the size we want, and sold later without giving the edge back.
The impact term is the one that is usually ignored, and it grows faster than size. There is always a notional beyond which acting on a correct view destroys its value. Below that point the view is an opportunity; above it, it is an anecdote.
| Question (illustrative) | Raw gap | Edge after costs | Capacity | Admissible |
|---|---|---|---|---|
| Policy path repricing by quarter-end | +11 | +8.9 | High | Yes |
| Volatility exceedance over one month | +13 | +4.1 | Moderate | Marginal |
| Regional credit spread threshold | +24 | +1.2 | Low | No |
| Commodity curve inversion | +19 | −0.6 | Very low | No |
The pattern in the table is general. Large gaps are disproportionately found in instruments that are hard to trade, because that is precisely where prices are slow to absorb information. A forecasting system that ignores execution will therefore drift towards the opportunities it is least able to capture, and will look most skilful exactly where it is least useful.
§ IIExecution changes the forecast
The less obvious argument runs in the other direction. When the Executor works an order, it observes things no forecasting model sees: how deep the book is at each level, how quickly liquidity replenishes, whether other participants appear to be trading the same view. That information is evidence.
If we try to buy protection that we believe is cheap and find that it is being bought aggressively by others, one of two things is true. Either the market is moving towards our view, in which case the gap is closing and the opportunity with it, or others have information we lack. Neither should leave our probability unchanged. Realised execution data is therefore fed back to the Forecaster agents as a feature, with a lag, and to the Auditor as a check on whether gaps that close quickly were ever real.
§ IIIProgrammable execution
Execution in the system is fully programmatic. The Allocator emits target exposures with constraints; the Executor decides how and when to reach them, subject to a cost budget derived from the edge it is trying to capture. A small edge earns a patient, passive execution. A large, decaying edge earns urgency. When the cost budget is exhausted before the target is reached, the remaining exposure is not pursued.
This last rule matters more than it sounds. It means that the system's exposure to any view is bounded not only by how strongly it holds the view, but by how much the market is willing to let it express. A system that always reaches its targets is a system that sometimes overpays to be right.
“A theoretical opportunity that cannot be executed is not an opportunity. It is a measurement error waiting to be discovered at size.”
§ IVOne process
From the outside, the stages of the system look sequential: observe, decompose, forecast, disagree, allocate, execute, learn. Internally, execution touches every one of them. It constrains which questions are worth asking, informs which gaps are real and supplies the data with which the system learns. We think of it less as the last step of the process than as the part of the process that meets the world.