Validation

No profitable strategy has been demonstrated and none is claimed. This page exists to say what would have to be true before anyone could claim one, and to record how little of that has been done.

Current status

The simulation runs, it commits its state, and the committed state replays. That is the whole of what has been established. It is a statement about plumbing, not about behaviour: it means the ledger is an honest record of what the model did, and says nothing at all about whether what the model did was any good.

None of the controls below have been run. Until they have, the correct reading of any run of good performance is that a five-asset allocation drifting on a random walk will sometimes drift the right way.

controls and baselines
TestStatusResult
Held-out chronological replaynot runnot yet measured
Frozen-weight controlnot runnot yet measured
Shuffled-price controlnot runnot yet measured
Shuffled-connectome controlnot runnot yet measured
Fees and slippage accountednot runnot yet measured
Compared against holding cashnot runnot yet measured
Compared against fixed equal weightsnot runnot yet measured

What would count as evidence

Evidence of learned behaviour is not a return figure. It is a return figure that survives every attempt to explain it away, and the attempts have to be made first.

Held-out chronological replay

Every parameter fixed on one span of market history, then run forward on a later span that nothing in the model has seen. Chronological, never a random split: shuffling time and sampling test days from the middle of the training period leaks the future backwards and will manufacture a result out of nothing.

Frozen-weight control

The same walk, with synaptic weights frozen at their initial values so no plasticity can operate. If performance does not fall, whatever the model is doing is not learning, and the plasticity is decoration.

Shuffled controls

Two of them, testing different claims. Shuffle the price series and keep the connectome: if performance survives, the result was never about the market. Shuffle the connectome — rewire it while preserving degree distribution — and keep the prices: if performance survives, the result was never about this animal, and the connectome is doing no work that a random graph of the same shape would not do. That second control is the one that matters most here, because the entire premise of the project is that the specific wiring is the point.

Costs

Every rebalance pays fees and slippage, and the fly rebalances on a walk, which is to say continuously. Costs have to be charged inside the simulation at the tick that incurs them, not subtracted from a total afterwards. A strategy that is profitable before costs and unprofitable after them is unprofitable.

Baselines

Against holding cash, and against fixed equal weights across the same five symbols, over the same period, with the same costs. Beating neither is the expected outcome and would be reported as such. A model that cannot beat a constant allocation has not earned the word strategy.

How results will be reported

Whatever comes out of the tests above goes on this page, including a flat negative — which is the most likely result and would be the more useful one, since it would say something about what connectome-derived dynamics do not give you for free.

Numbers will appear here only once they have been produced by running the thing they describe. Where a value has not been measured, this page says not yet measured rather than an estimate, a projection, or a figure from a related system. Nothing here is investment advice.

The connectome is anatomy. Sensory encoding, membrane dynamics and the mapping from body position to portfolio weight are engineering choices, described in the model and the arena, and any of them could be wrong.