Our research standard
Hypothesis first. Validation always.
This is the method every idea at Zeviq goes through, whether a person or an agent proposed it. It is a statement of how we work, not a report of results.
Eight rules
Every idea is a hypothesis with a named mechanism.
Before anything is tested, we write down why it should work: who is on the other side of the trade, and why they would keep being there.
Pass criteria and falsification come first.
What would count as a pass, and what would prove the idea wrong, are written before the data is touched. The bar does not move after the results are in.
Out-of-sample, always.
Walk-forward validation and held-out periods the idea has never seen. Results are reported with bootstrap confidence intervals, not single point estimates.
Multiple testing is counted.
Every variation tried is recorded, including the ones that failed. Sharpe ratios are deflated for the number of trials, so a lucky draw from many attempts is not mistaken for an edge.
Costs and realistic fills in every backtest.
Commissions, spread, and slippage are always on. Fills are modeled conservatively: for example, a passive order only fills when price trades through it.
The builder never validates their own result.
Independent reviewer agents and people check the work, the code, and the conclusion before anything moves forward.
Fat tails are assumed.
Market returns are not normally distributed. Features are built on returns, not prices, and risk is sized for bad paths, not for the average one.
Agents supervise, code executes.
Agents research, review, and monitor. Orders and risk limits belong in deterministic, tested code that an agent can halt but never loosen.