SPX / SPY directional edge · updated daily
The market overpays for SPX crash protection. We sell it — carefully.
Over 1993–2026, the option market persistently overprices SPY's downside — the volatility skew charges a premium for puts that expire worthless ~90% of the time. Sold as a weekly ladder of defined-risk put credit spreads under a fully audited pricing model (carry, skew, mean-reverting IV, real slippage), that edge compounds at ~25–30%/yr at roughly half the index's drawdown.
Loading latest signal…
Today's positions
Two books, opposite ends of the frontier
Both harvest the SPY risk premium as ladders, entered only in an uptrend. The strategy is the put credit spread: sell −3% / buy −6%, ~3 months out, held to expiry, a new rung every week (3% of equity per rung, 60% at-risk cap) — it gets paid by the volatility skew. The call spread (+2%/+7%, 1 year, GTC at 80% of width, monthly) is the max-ROR-per-trade alternative.
Loading…
Anatomy of one rung, in dollars
What this week's rung looks like per contract, and the account size for which one contract equals the reference sizing (3% of equity at risk per rung).
Loading…
Cadence
How often it trades, and what triggers a trade
There is no prediction and no signal to wait for. A trade is triggered by the calendar plus one filter: on the first session of each week, if SPY closed at or above its 200-day average, sell the next rung. That's the entire trigger. Exits need no decision at all — every rung simply expires ~13 weeks after it was opened.
A typical week takes about five minutes: check whether SPY is above its 200-day average (the page header shows this). If yes, sell one put spread — strikes 3% and 6% below spot, ~3 months out, sized to 3% of equity at risk — and place no other orders. If no, place nothing and let the open rungs run off. Any rung expiring this week settles by itself. When the regime turns off mid-bear, the book winds itself down to cash within ~13 weeks without a single sell decision.
Track record
Validated 1993–2026, out-of-sample from 2016
Every eligible entry day, held under the exact live exit rule (thousands of samples). Win rate and ROR are the robust full-sample figures; CAGR and drawdown come from the ladder books (weekly rungs for the put book, monthly for the call book, exits applied as they occur).
| Book | Win rate | Out-of-sample | Mean ROR | Median ROR | Avg hold | Annualized | Worst trade |
|---|---|---|---|---|---|---|---|
| Loading… |
Equity & risk
The strategy equity curve, 1993–2026
The bold navy line is the strategy — the weekly put-spread ladder, risking 3% of equity per rung with a 60% total at-risk cap, priced under the audited v2 model. Lighter lines: the call ladder and SPY buy-and-hold. All start at 1.0; log scale. Weekly laddering keeps ~10 rungs working at once — ~27%/yr at roughly half the index's drawdown (design era 25%, untouched 2016–2026 validation 30%).
Drawdown
How far each book sat below its own prior peak — the lived experience of holding it.
Distribution of per-trade returns
Where the ROR of each closed trade lands. The put book banks its credit ~90% of the time; the losers are bear-market breaches, capped by the long leg.
Example trades
A representative winner and the worst loser from the put book — the shape of both outcomes.
Where the edge comes from
A real disagreement, not a forecast trick
The market overprices the downside
The volatility skew makes SPY puts the most systematically overpriced options listed: a −3% put carries a higher implied vol than any call, yet expires worthless ~90% of the time in an uptrend. Selling a defined-risk put spread collects that premium with the loss capped by the long leg — the seller is paid by the skew, twice over when entries are gated to uptrends.
Options convert the edge to ROR
Holding SPY earns the drift (~10%/yr). A defined-risk credit spread is leveraged: each rung risks a small, fixed amount to collect the skew premium — +21% mean ROR per ~3-month trade at an 88% win rate, recycled weekly.
Hold to expiry — exits give back the edge
We tested every exit: profit-capture buybacks, stop-losses, regime-break dumps. Under honest pricing they all pay extra slippage to give back the tail of the premium. The put book holds each rung the full ~3 months and lets the skew decay to zero; only the call book keeps a GTC limit (its first-passage profit is real).
A regime filter, not a hedge
The losers cluster in bear markets, so the book only opens new trades when SPY is above its 200-day average and stands aside in downtrends. This free filter beats every hedge we tested: it lifts win rate 81→87% and cuts the drawdown 39→28%. Buying puts to hedge does the opposite — it pays away the very premium the strategy harvests and erases the edge.
Honest about the frontier
This is ~88–93% accurate, not 99%. The ~1-in-9 losers cluster in bear markets, and concurrent rungs lose together in a crash — the 60% at-risk cap is what bounds that. Position sizing, not the signal, sets your drawdown.
What "ROR" means and doesn't. Turning per-trade ROR into portfolio CAGR is a sizing choice on the ladder: 2% of equity per rung compounds ~19%/yr at a −23% max drawdown; 3% (the reference) ~27%/yr at −31%; 4.5% ~35%/yr at −41%. The edge is real; the leverage is yours to set.
The pricing model is audited, not convenient. Backtest premiums include the cost of carry (historical T-bill rates), the real volatility skew (OTM puts priced rich, OTM calls cheap), mean-reverting long-dated IV (2008 prices at ~46%, not 90%), and 3% slippage each way. These corrections cut the naive backtest's numbers by more than half — what remains survives every perturbation we threw at it (steeper skew, cheaper IV, 5% slippage: 19–22%/yr). SPY/SPX options are the most liquid listed, so the modeled fills are attainable.
The rulebook
Exactly which options we sell, and why — step by step
There is no black box. The whole strategy is five mechanical rules, each chosen by measurement on the design era (<2016) and confirmed on untouched 2016–2026 data. Here is the complete logic, the edge behind it, and how every premium is priced.
The weekly decision
| Step | Rule | Why — and the evidence |
|---|---|---|
| 1 · Check the regime | Is SPY at or above its 200-day average? If not, do nothing this week. | Put-selling losses live almost entirely in established downtrends. Adding this one filter lifted the win rate ~6 points and cut the max drawdown by a quarter — the only "free" improvement found. Every hedge tested (buying puts, stop-losses) cost more than it saved. |
| 2 · Pick the strikes | Sell the put 3% below spot; buy the put 6% below spot. | −3% is the sweet spot where the skew premium is rich but breaches are still rare (see the edge numbers below). The −6% long leg caps the worst case and defines the risk. Grid-tested against −5%/−10%, −4%/−9%, wider and narrower — on design-era data only. |
| 3 · Pick the expiry | ~3 months out (63 trading sessions). | Short enough that the premium decays fast and capital recycles ~4×/year; long enough to survive noise. Tested against 1-year (carry drag, slower recycling) and 1-month — 63 sessions won on design data and validated better out-of-sample. |
| 4 · Size the rung | Risk 3% of current equity; stop adding rungs at 60% total at risk. | Sizing is the only real risk dial: 2%/rung ≈ 19%/yr at −23% max drawdown, 3% ≈ 27%/yr at −31%, 4.5% ≈ 35%/yr at −41%. The cap bounds how much a crash can take — all open rungs lose together in a bear. |
| 5 · Hold to expiry | No profit-taking, no stop-loss, no adjustment. Let it expire; enter the next rung next week. | Every exit tested gives back edge: profit-capture buybacks pay slippage to surrender the tail of the decay, stop-losses whipsaw, regime-break dumps sell the lows. The premium is earned by carrying the position to zero. |
The edge, in two numbers
This is the entire alpha. Option prices imply a probability that SPY falls through the −3% strike within 3 months. Reality, measured across every uptrend entry since 1993 (6,215 samples), is far kinder to the seller:
Put differently: the buyer of this spread is buying insurance, and like most insurance it is priced well above actuarial cost. Who pays it, and why it persists: institutions are structurally obliged to hedge (mandates, regulation, career risk), so they buy index puts at almost any price; the skew premium is their willingness-to-pay. The seller earns it by bearing crash risk — which is exactly why the regime filter, the −6% long leg, and the at-risk cap exist. The premium has persisted for three decades because the occasional crash keeps most sellers away.
How every option is priced — today's rung, leg by leg
The backtest never uses a magic price. Every leg is priced with Black-76 on the forward, with the volatility read off an audited surface. This is today's actual computation, live from the signal:
Loading…
Every modeling assumption, and its evidence
| Assumption | Value used | Evidence it's right (or conservative) |
|---|---|---|
| Pricing formula | Black-76 on the forward: F = S·erT; puts valued at e−rT[K·N(−d₂) − F·N(−d₁)] | Standard European options pricing. The panel is total-return SPY, so the self-consistent carry is the full T-bill rate (dividends already live in the path). |
| Interest rate r | Historical 3-month T-bill, by year (5.5% in 1995 → 0.05% in 2021 → ~4% now) | Setting r=0 (the naive choice) overstated the strategy's CAGR by ~8 points — the single biggest error found in the pricing audit. Corrected. |
| ATM implied volatility | 1.15 × √(0.30·rv₆₀² + 0.70·r̄v²) — a mean-reverting blend of current 60-day realized vol and its expanding historical mean | Produces a mean 1y ATM IV of 18.6% (real SPX ≈ 19%), a 2008 peak of ~46% (real long-dated IV peaked ~45–50%, while the naive proxy said an absurd ~90%), and dead-calm ~16%. See the IV chart below. |
| Volatility skew | IV(K) = ATM · (1 + ln(F/K)) — each leg gets its own vol: today the −3% put prices ~0.7pt over ATM, the −6% put ~1.3pt over | Matches the typical SPX skew slope. Crucially this makes the model charge us more to buy the −6% leg — the conservative direction. Results hold at double the skew slope (β=1.8 → still 19%/yr). |
| Slippage | 3% of spread value each way (≈17¢ round trip on today's ~$5.70 spread) | SPY options are penny-to-nickel wide at these strikes — the most liquid options listed. 3% is generous; at 5% the strategy still makes 21%/yr. |
| Fills | Enter at model mid − slippage; expire at intrinsic value | No intraday timing is assumed anywhere — entries use the daily close, expiries settle at intrinsic. Nothing depends on execution skill. |
What the model deliberately does not capture — disclosed, not hidden: SPY options are American-style, so a short put that goes deep in-the-money can be assigned early (worst case unchanged — you'd own stock hedged by the long put, same defined loss, but know it can happen). Real strikes come in $1 increments and real expiries on fixed Fridays, so live strikes/dates will sit within ~½% of the model's. And all premiums are modeled, not historical quotes — the audit (below) is our best effort to make the model honest, and every number survives adverse perturbations of it.
Checked against live option prices — SPY and SPX
Every night the model is put on trial against the real market: we pull the live SPY and SPX option chains, find the actual listed contracts closest to the strategy's legs (−3% / −6%, ~91 days), and compare the credit the model books against what the market genuinely pays — at mid, and at the natural worst-case fill (sell at the bid, buy at the ask). The model passes if it books no more than the worst-case real fill.
First live check runs tonight after the close.
What the first live check found — reported straight. In today's unusually calm market (multi-year-low volatility), the model books more credit than the market currently pays: real mid ≈ 18% of spread width vs the model's ~25%. Two offsetting facts: real liquidity is far better than modeled (bid-ask 0.3–1.4% of mid vs our 3% slippage assumption; thousands of open contracts), and the gap reflects the calm regime — the model's IV blend runs rich when volatility is at cycle lows and steep skew makes the protective leg relatively dear.
So we stress-tested it: haircut every entry credit in the entire 33-year backtest to today's lean level (×0.70) and re-run the ladder. Result: 11.4% CAGR at −32% drawdown (design 10.8%, validation 12.5%, win rate unchanged at 88%) — the strategy still beats SPY buy-and-hold (10.8% at −55%) even if every week forever paid like today. The honest expectation is the bracket: ~11%/yr if premiums are always this lean → ~27%/yr at modeled premiums, with the perturbation tests (19–22%) in between. The nightly check above keeps score.
The same strategy on SPX index options is structurally better. SPX options are European-style (no early assignment — that caveat above disappears), cash-settled (no stock delivery, positions simply settle to dollars at expiry), typically taxed 60/40 long/short-term in the US (Section 1256), and 10× SPY's size (one SPX spread ≈ ten SPY spreads — fewer contracts, less commission for larger accounts; XSP is the SPY-sized mini with the same European cash-settled terms). The skew premium being sold is the same S&P 500 downside surface, so the strategy ports unchanged: same strikes in percent, same tenor, same cadence, same sizing.
Downside protection
What happens in bears — including ones that never happened
The protection is structural, not a hedge: every rung's loss is capped by the long put (defined risk); total exposure is capped at 60% of equity; and the 200-day filter stops new rungs in a downtrend, so the book is fully flat within ~3 months of a regime break and stays in cash while the market keeps falling. We tested it against every historical bear and synthetic bears built to be worse than anything since 1993.
Every historical bear hurt the strategy less than holding SPY
| Drawdown window | Strategy | SPY buy & hold |
|---|---|---|
| Dot-com bear · Sep 2000 → Oct 2002 | −27.6% | −47.3% |
| Global financial crisis · Oct 2007 → Mar 2009 | −17.8% | −55.2% |
| COVID crash · Feb 2020 → Mar 2020 | −6.0% | −33.7% |
| 2022 bear · Jan 2022 → Oct 2022 | −23.4% | −24.5% |
Synthetic stress: prolonged bears with no historical precedent
Paths block-bootstrapped from real SPY returns (volatility
clustering preserved), spliced onto real history so the regime and volatility state are
realistic at stress onset, then re-drifted into shapes SPY has never produced.
Reproduce with strategies/spx_predict/stress.py.
| Scenario | Strategy | SPY buy & hold |
|---|---|---|
| Japan-style: −60% crash, then 7 years flat | +54.0% | −61.0% |
| Lost decade: −5%/yr grind for 10 years (seed 1) | −17.9% | −40.1% |
| Lost decade (seed 2) | −37.4% | −40.6% |
| Lost decade (seed 3) | +8.2% | −39.8% |
| Whipsaw grind: −20% legs + 16% re-arming rallies, 8 years | −94.4% | −56.4% |
Crash-then-stagnation is where the strategy shines most. It doesn't need the market to rise — only to not fall 3%+ over each ~3-month window. In the Japan path it made +54% while the index lost 61%: after the crash the filter kept it flat, and through the long stagnation it kept collecting put premium in a market going nowhere. In grinding −5%/yr declines it beat buy-and-hold in 3 of 3 seeds.
The honest failure shape: the whipsaw. A path of repeated −20% legs, each followed by a +16% rally that recrosses the 200-day average, re-arms entries at every local top — and each new leg kills the fresh rungs. On a path built to do exactly that for 8 years, the strategy loses −94% vs −56% for buy-and-hold. Nothing like it exists in post-1993 SPY (1929–42 and 1968–82 partially rhyme). The mitigation is sizing, not signal: halving the rung to 1.5% roughly halves every stress number. We disclose this because a strategy whose worst case you haven't seen is a strategy you don't understand.
Methodology & research log
How this strategy was actually built
The full research trail, in the order it happened — including the findings that killed earlier versions. Every number is reproducible from the source; the complete write-up is in VALIDATION.md.
1 · Data & discipline
SPY daily (total-return adjusted), 1993–2026, ~8,400 sessions. Everything is point-in-time: any probability or indicator at date t uses only data before t. Structures and parameters were selected on the design era (<2016) only; 2016–2026 is untouched validation. Where the validation decade outperforms the design era — as it does for the final strategy — that is evidence the edge is durable rather than fitted.
2 · The measured edge
The option market prices SPY near risk-neutral; the physical distribution drifts up and crashes less than implied. Actual frequency above spot beats option-implied probability by +9pp at 5 days → +35pp at 1 year. The volatility skew concentrates that mispricing in downside strikes: OTM puts are the most systematically overpriced options listed. That premium — not a forecast — is what the strategy collects.
3 · Findings that shaped (and killed) earlier versions
| Question | Measured answer |
|---|---|
| Can we be 99% accurate AND disagree with the market? | No — calibration saturates at ~98% (crash onsets are unforecastable); at 99%+ the market already agrees, so the edge is ~0. The accuracy/edge frontier is win ≈ 1/(1+payoff). |
| Close early to lock profits? | Lifts win-rate to ~99% but collapses CAGR below buy-and-hold (avg loss ≈ 20× avg win). Negative skew, not edge. |
| Stop-losses? | Whipsaw. Tight stops cut the win rate to 60%; loose ones are CAGR-neutral and can't stop gaps. |
| Buying protective puts? | Self-defeating — pays away the exact premium the strategy harvests; CAGR → ~0%. |
| Regime filter (200-day average)? | The one free improvement: lifts win rate, ROR, and drawdown simultaneously by standing aside in downtrends. |
| One position at a time? | Wastes capital — laddering entries (weekly rungs, at-risk cap) roughly doubles CAGR at the same drawdown. Entry day-of-month is irrelevant. |
4 · The pricing audit that inverted the strategy
Early versions priced options with flat IV (60-day realized × 1.12) and zero interest. Auditing that model against real market features found four material errors. Each is toggleable in the code; the attribution on the then-champion call-spread ladder:
| Pricing correction | Why it matters | CAGR impact |
|---|---|---|
| Cost of carry (historical T-bill forward) | The panel is total-return SPY; r=0 overstated the edge by the bill rate, leveraged | 28.1% → 20.5% |
| Volatility skew (β=1.0) | OTM calls actually trade cheap, OTM puts rich — v1 sold the call leg at fantasy vol | 28.1% → 19.6% |
| Mean-reverting blended IV | 1-year IV priced 2008 at ~46%, not ~100%; sane calm-market levels (~16%) | 28.1% → 24.8% |
| 3% slippage per side | Long-dated spreads are wider than fronts | 28.1% → 26.7% |
| All corrections together | debit/width 41% → 54% (the realistic figure) | 28.1% → 10.0% |
Under honest pricing the frontier inverts: buying call spreads pays the carry and buys the expensive side of the skew; selling put spreads collects both. Re-selected on the design era only and confirmed out-of-sample:
| Book (ladder, v2 pricing) | Design <2016 | Validation 2016–26 |
|---|---|---|
| Put credit spread −3%/−6%, 63d, hold to expiry, weekly | 25.2% / −31% | 30.3% / −28% |
| Call spread +2%/+7%, 252d, GTC 80%, monthly | 9.9% / −30% | 13.5% / −26% |
5 · Robustness of the selected strategy
The put ladder was perturbed on every pricing assumption (full period, 1993–2026): steeper skew β=1.4 → 20.5% CAGR; β=1.8 → 19.1%; cheaper IV (×1.05) → 19.2%; spikier vol blend → 21.3%; 5% slippage → 21.0%. Win rate pinned at 88% throughout. The edge is not an artifact of any single modeling choice. Entry cadence is likewise insensitive (weekly/biweekly/monthly all work; weekly is smoothest).
6 · Example trades from every market era
A representative winner and the worst loser from each era of the backtest — entry, strikes, premium collected, capital at risk, and outcome. Dollar figures are per contract (×100 shares).
| Era | Entered | Sold / bought strikes | Credit | At risk | Exited | Hold | ROR |
|---|---|---|---|---|---|---|---|
| Loading… |
7 · Honest limits
Premiums are modeled (audited v2 surface), not historical quotes — SPY/SPX options are the most liquid listed, so modeled fills are attainable, but they are still a model. The strategy is ~88% accurate, not 99%. Losers cluster: concurrent rungs fail together in a crash (bounded by the 60% cap). The equity curve marks at exits; intra-month marks sit lower. And the whipsaw failure shape above is real, if unprecedented. What we will not do is present a backtest we couldn't defend trade-by-trade.
Recent closed trades
Every rung the ladder took, most recent first
All closed ladder rungs, newest first. "limit"/"profit" = closed early on the GTC target; "expiry" = held the full year.
| Book | Entered | Exited | Strikes | Hold | Exit | ROR |
|---|---|---|---|---|---|---|
| Loading… |