CreditFloor hunts for (stock, expiry, strike) triples where every one of thousands of historical out-of-sample tests closed a winner. The engine runs in both directions: a put-credit floor below spot where the stock's close never once fell below the strike before expiry, and a call-credit ceiling above spot where the close never once rose above it. We're deliberately selective: most stocks, most days, there’s no trade. When the engine does publish, it publishes a lot: the exact strike, the exact expiry, the fold-by-fold walk-forward breakdown, and the worst-case buffer ever touched on the test set. Not financial advice.
Every cron run appends every (ticker, side, horizon) signal it publishes to an immutable log. When a signal's expiry date arrives, we look up the actual close on that day and mark the outcome — win if the close is on the safe side of the strike, loss otherwise. Signals for tickers that later drop off today’s list stay in the log and still resolve at their expiry: this is the anti-survivorship machinery. A signal published on the day a ticker dropped off is still held through its full horizon and graded honestly.
Each card is a ticker whose walk-forward, cross-fold, out-of-sample win rate is exactly 100%. Multiple expiries ("rungs") may be eligible; each rung shows its own strike and buffer. Smaller buffer = strike closer to spot = richer credit with the same historical guarantee. Sections below start collapsed — tap to expand.
The highest-conviction opportunities from today’s scan — pulled from both the put and call sides, filtered to rungs with estimated per-trade return-on-risk at 5% or higher, and sorted biggest to smallest. Every one of these rungs has a 100% walk-forward out-of-sample win rate. Credit, max loss, and ROR are Black-Scholes estimates (r=0, IV = realized × 1.30, 20% haircut for bid-ask, 5%-of-spot spread width) — real market fills will differ.
A separate strategy from the credit floors/ceilings above. These spreads fire only when a stock enters a specific oversold/overbought regime (SPY-relative weakness, multi-stack mean-reversion, etc.), targeting short-dated (5-21 day) credit spreads with a fixed 3% spread-width. Win condition: stock's close at expiry stays on the safe side of the short strike. Historical accuracy 83-91% per rule, tested on 94 liquid-options names. All three sections below start collapsed; tap to expand.
Iron-condor credit spreads with proprietary 6-layer robustness (two-stage walk-forward, stress-pricing eligibility, regime σ envelope, immutable live log) covering all four corners of the {WR, ROR, liquidity, frequency} Pareto. Pick the tier that matches your tradeoff. Each section starts collapsed; tap to expand. Live track record below shows live vs backtest WR per tier.
At time t, the stock is in regime iff every
one of these causal features clears its threshold:
vol20 < 40% (annualized stdev of log returns).vol5 / vol20 < 1.05 — recent vol at or below trailing vol.< 15% of spot.|close / SMA20 − 1| < 4%.RSI14 ∈ [25, 75].|5d return| < 8%.
Conditioned on this regime, the empirical distribution of next
h-day path moves (for h ∈ {5, 7, 10, 14, 21})
has a 97th-percentile materially smaller than the unconditional
distribution. That is the entire edge.
Core tier (static conformal) — for each (ticker, side, horizon ∈ {5,7,10,14,21}):
b̂core = quantile0.97(b∗ on training) + 0.5%
By the order-statistic interpretation, this targets ~97% in-sample coverage. Capped at 5% buffer.
Tight tier (vol-adaptive conformal, novel) — for each (ticker, side, horizon ∈ {2,3,5}):
b̂tight = z0.95 × σtoday × √(h/252) + 0.5%
where z = b∗ / (σ × √(h/252))
is the path buffer normalized by an instantaneous vol estimate
(vol20) at each historical day. The
quantile is taken on these normalized values, then re-scaled by
today's σ at publish time. When current vol is below
regime-historical levels, this collapses the buffer to a fraction
of the static-quantile floor — producing strikes
as tight as 0.86% from spot on 2-day expiries
while preserving 95%+ accuracy. Capped at 3% buffer; uses a
stricter "tight" gate (vol20 < 30%,
compression < 0.95, range < 10%, flat < 2.5%, RSI in
[35, 65], |5d ret| < 5%).
(t, t+h] closes strictly before Jan 1 of
the test year — purged gap, no overlap.b∗(t,h) ≤ b̂.b̂ ≤ 5% (core) / 3% (tight).When BOTH sides clear all gates simultaneously, the resulting short-put + short-call combination forms a tight iron-condor-equivalent pin trade: spot is expected to "pin" within a narrow band through expiry. Both sides each carry their own ≥95% historical win rate; the combined credit can offset two-way risk.
The two single-side tiers max out at ~15% per-trade ROR
because credit-spread math constrains it: short-strike
delta is essentially 1 − WR, so 95% WR
implies ~5% delta which means thin credit. Iron
condors break this constraint by stacking credits:
put-breach and call-breach are mutually exclusive at expiry,
so combined credit is both legs while max-loss is
still bounded by one width minus combined credit. Per-leg
ROR of 20% becomes combined ROR of 50%+.
Joint walk-forward validation: a fold-day counts as a win
only if BOTH the put-side path-min stays
above Kput,short AND the call-side
path-max stays below Kcall,short.
Per-leg conformal at quantile q ∈ {0.97, 0.975, 0.98}
— the engine picks the smallest passing q per
ticker (tightest strikes that still validate). Spread width
is also swept across {1%, 2%, 3%, 5%} of spot
and the engine keeps the width that maximizes combined ROR.
Walk-forward 95%+ OOS accuracy is empirical; future paths can break the conditional bound. The regime gate is the entire defense — if a stock is no longer in stillness today, we do not publish. Position-size as you would any short-vol strategy.
For every historical day t and every horizon
h ∈ {21, 42, 63, 126} trading days we compute two
path statistics:
b∗⊂put(t, h) = 1 − min(close[t+1..t+h]) / close[t] — how far below the entry price the stock ever traveled.b∗⊂call(t, h) = max(close[t+1..t+h]) / close[t] − 1 — how far above the entry price it ever traveled.
A put-credit spread with short strike K = close[t] × (1 − b)
is safe iff b ≥ b∗⊂put; a call-credit spread with
short strike K = close[t] × (1 + b) is safe iff
b ≥ b∗⊂call. Both are strictly tighter than
"close inside the strike at expiry" — a passing signal also
covers American-style early-exercise risk.
For each fold year Y ∈ {2020…2026} we hold
that calendar year out as test. Training uses only samples whose
forward window closed before Jan 1 of year Y
(a purged gap equal to the horizon — no overlap with test
dates). Our conformal strike buffer is
b̂ = max(b∗ on train) + 1% — the largest
buffer ever needed in training, plus a flat safety margin.
By the order-statistic bound this has 100% empirical coverage on
training; we then measure how often it also covers the test year.
The method and the 1% safety margin are identical for both sides.
We run each ticker twice on each side: unrestricted, and restricted to a side-specific regime.
close ≥ SMA200 AND 252-day drawdown ≤ 20%.close ≤ SMA200 AND ≤ 20% rally off the 252-day low.The regime-gated variant is only deployable when today’s features also satisfy the gate. Per ticker we keep whichever variant produces a smaller final buffer — never the one with a better win rate, because both must already be 100%.
A (ticker, horizon) combo is eligible only if every walk-forward fold has a 100% win rate with ≥1 test sample, at least 50 pooled OOS tests, and the final live buffer is ≤25%. After per-ticker filtering we pool every remaining OOS prediction across the entire universe and require the aggregate win rate to still be 100%. If a single historical trade failed, the entire signal is rejected.
t use only close[0..t]; RSI, SMA, volatility, drawdown, momentum all causal.(t, t+h]; never mixed into features.A 100% historical OOS win rate does not promise the future. Regime shifts, a bigger-than-2020 crash, or a ticker-specific catastrophe can all breach any strike. Position-size accordingly, trade as a portfolio, and assume future drawdowns can exceed the worst historical one we trained on.
This page is a quantitative research artifact, not financial advice. Past performance, including the walk-forward out-of-sample results shown here, does not guarantee future results. Options trading involves substantial risk, including the possibility of total loss of premium collected and losses beyond the premium if the spread is breached. The short-strike prices published are derived from the empirical worst historical drawdown plus a fixed safety margin; future drawdowns can exceed anything observed in training or validation. Consult a licensed financial advisor before trading.