A backtest is a hypothesis, not a license to trade
A backtest alone is not evidence that a strategy works. Here, a good backtest earns a strategy the right to be forward-tested on paper — and nothing more — until live, unedited results confirm it.
This post is the live status of that process. No cherry-picking, no “trust us.”
The one edge in validation: VRP on OKX ETH options
After proving directional prediction does not beat fees, one non-directional edge survived a beta-strip: selling volatility (the volatility risk premium). It was a strong lead. The gate below has since run, and it failed — here is exactly where it stands:
Verdict published (2026-07-22, OOS-confirmed 2026-07-23): NO-GO. On real Deribit bid/ask snapshots a 7DTE iron condor earns a negative Sharpe before cost (ATM fly −0.09, 1SD condor −0.29) and −0.49 / −0.68 after spread and fees (
research/findings/20260722_vrp_condor_realfills.md:3-8). The strategy is markedkilled. This page said “the backtest is running, the verdict will be published here” for six weeks after the verdict existed — the lag is on us, and this notice is the correction (2026-09-06).
- What it is: weekly defined-risk iron condors on OKX ETH options. One contract is 0.1 ETH — about $246 of notional at ETH $2,457 (OKX instruments API, 2026-09-05) — so a condor capped at about 10% of notional risks about $25, roughly 0.5% of a $5k account. Small enough to run OKX-native, no other venue. Corrected 2026-09-05: previously
~$1,700 of notional … ~2–3% of a small account, which counted a whole ETH as one contract. Dollar amounts are a snapshot, not auto-updated. - Why it is not live, and will not be: the edge was proven on a proxy volatility model, and the proxy was the problem. Re-derived with a true condor payoff and realistic cost, the proxy’s Sharpe 2.0 becomes 0.92 (
research/findings/vrp_audit_math_20260620.md:5,25); on real quotes it goes negative. A 2026-08-01 pre-screen of a two-leg strangle as a substitute vehicle is still not a GO (research/findings/strangle_cost_prescreen_20260801.md:3). - The gate did its job: it was “the condor backtest on real bid/ask must hold.” It did not hold, so there is zero real capital and no phase-in. That is the gate working, not the gate pending. This bullet used to read “Update (2026-08-14): sufficient depth reached, the full real-chain condor backtest is now running.”
Paper trackers running now
Alongside the data collection, we run forward paper trackers that take the real signals in real time and record what they would have done — with no benefit of hindsight:
- A combined VRP + trend tracker (the two are uncorrelated, so together they diversify).
- A trend-ensemble tracker (managed beta — risk control, not alpha).
These exist to catch the gap between a clean backtest and messy reality before money is on the line. If the paper results diverge from the backtest, the strategy does not graduate. Several already have not — and we keep them visible as failures rather than quietly deleting them.
Why we publish the waiting
It would be easy to announce “we found a winning strategy” today. We won’t, because the honest answer is “we found a strong lead, and it has to survive a real-data gate first.” That discipline — paper before real, real-chain before proxy, phases before full size — is the entire difference between research and gambling.
When the gate passes, you will see it here with the real numbers. When something fails the gate, you will see that too. That is the deal: we show what fails, we show what is still being tested, and we only call something live when it actually is.