06 Forecasting · Market microstructure · Measurement
Financial forecasting
Instruments that say when an edge isn’t there.
The question
Can a small research loop tell a real trading or forecasting edge from a lucky one before real money finds out?
- Role
- Research direction; designed and ran the instruments
- Years
- 2026
- With
- Independent
- Status
- Research
- Disciplines
- Forecasting, Market microstructure, Measurement
- Constraint
- Markets flatter confident backtests: gaps in the data, optimistic fills and a model that can read the very market it’s scored against all manufacture edges that aren’t there.
Method
01/05
Order books, prices and funding are recorded to a local database, and new ideas run on paper against the recordings before anything else.
Demo data, not trading results
02/05
Each idea goes into a research log with its test and its verdict, and is judged out of sample with pessimistic costs.
Demo data, not trading results
03/05
Profits the order log can’t support are thrown out; every number on the dashboard has to be defensible by hand.
Demo data, not trading results
04/05
A harness around an open-source forecasting bot logs its forecasts beside prediction-market prices, with its own market research switched off so it can’t read the answer.
Demo data, not trading results
05/05
Kill switches, budgets and read-only checks stand between any strategy and a live order.
Demo data, not trading results
Results
- No deployable edgeOvernight lab: backtest verdicts on spot-crypto strategiesBacktests and paper trading; no live results are cited.The overnight lab’s verdict on its own spot-crypto backtests: the shipped momentum strategy lost out of sample even at zero fees.Backtests; no live results.
- 37.7% worseKalshi research log, RL-003: can LLMs price markets better than the market?Phase 2 calibration: a four-agent open-source LLM ensemble and a single-model baseline, scored by Brier score against the market price on 45 of 50 selected markets. The expectation was recorded after the test.An open-model LLM ensemble’s Brier score against the market price, on 45 of 50 selected prediction markets — ruled out as a way to price them.A calibration test, not trading.
- Phantom profit, removedHyperfluid profit-fix plan, Phase 0: make the numbers trustworthyPaper trading. The audit replaced a dashboard alert’s funding profits with the net paper result the order log supports.An audit of a funding-rate carry book replaced its dashboard’s reported profits with the net paper result the order log supported.Paper trading.
Carried forward
The most valuable thing a trading system can give you is a clear no.
Credits
- Ryan BrantResearch direction; designed and ran the instruments
- Jan F-F and contributorsThe open-source forecasting bot (MIT) the harness extends