Research log
What passed, what failed, and why.
A selection of recent results from the KynovaX platform, in the order a reader might ask about them. Numbers are research results after trading costs. Sharpe is a standard measure of return per unit of risk: above 1 is good, around 0 means no edge.
Read this first. These are backtests on historical data and paper trading on a demo account. They are not investment advice, not an offer, and not a promise of future results. Backtested results have limits.
| Idea | Result | What we found | Tested on |
|---|---|---|---|
| Bitcoin trend breakout (4-hour bars) | PASSED | Passed all five gates and the skill check. On hidden data (Nov 2023 to Jun 2026) it made a Sharpe of 0.97 after costs over 79 trades, with timing beating random timing (p = 0.055). Now paper-trading; the forward record is far too short to judge. | Bitcoin, hidden data |
| Bitcoin overnight rule (published paper) | FAILED HIDDEN DATA | Buy at a 10-day high at the US close, sell the next morning. In research: Sharpe 1.23 against 0.70 for buy-and-hold, timing far better than random. On hidden data: Sharpe 0.00 over 89 trades, timing no better than random. It did not persist, and we did not tweak and retry. | Bitcoin, 2017 to 2026 |
| Eight academic papers, replicated | NONE PASSED | We read the exact rules of eight finance papers and tested each on our data with real trading costs. One looked strong in research (the overnight rule above) and then failed. Intraday gold and crude oil patterns, noise-area breakouts and cross-asset momentum did not survive costs or fresh data. | Gold, oil, indices, FX, crypto |
| Momentum filters for breakout strategies | NO BETTER THAN CHANCE | Nine popular filters (ADX, efficiency ratio, trend, 12-month momentum and others) were tested on four breakout types across nine markets, about 41,000 trades. None beat randomly removing the same share of trades. A quiet-market filter was closest and helped on some markets. | 9 markets, 30-minute bars |
| Central bank announcement reversal | NOT PASSED | A 2026 paper reported that stock moves on US interest-rate decision days partly reverse over the following month. The slope had the paper's sign, but trading costs and financing absorbed it (average trade -0.22%, not significant). | S&P 500, Nasdaq, Dow |
| "Sell in May" seasonal rule | NOT PASSED | Winter returns were higher than summer by a gap too small to rule out luck, and holding stocks only half the year earned a worse return per unit of risk than simply holding them on all nine indices. | 9 stock indices |
| "Buy the fear spike" (VIX above 30) | NOT PASSED | Failed the luck-adjusted gate on all nine indices. A side finding: our Nasdaq strategies earned more when the VIX was below 25 than above 20. | 9 stock indices |
Selected from the platform's full record of about 7,000 strategy tests. Figures as of 8 October 2026. We pre-register each test: the rule and the pass mark are committed before the result is known.
How to read a result
- Gates. Five fixed checks every strategy must pass: average Sharpe, Sharpe adjusted for how many tries we have made, stability across nearby settings, positive results in most 2-year blocks, and timing that beats random entry dates.
- Skill check. Separates a strategy that earns more than its own market from one that only rides a rising market.
- Hidden data. A locked slice of history, opened once per strategy. A pass here is much stronger evidence than a pass in research. Even then, it is evidence, not proof.
- Demo account. Passing strategies trade automatically on a broker demo account at 0.1% risk per trade with a stop on every trade, so we see how they behave going forward.