Saty's weekly ATR level probabilities are stable across decades. We priced every level against real SPXW option quotes for six years and found exactly one spot the market pays too much for: puts at the −1 ATR level. One simple weekly credit spread harvests it.
Saty Mahajan's ATR Levels project a set of price rails from two inputs: the previous period's close, and the previous period's 14-period Wilder ATR. On the weekly timeframe, the level at prior close − 1.0 × ATR has been crossed at Friday's close in about 5–6% of weeks — and that number barely moves across 26 years of SPY data, across the 2020–2026 SPX sample, and across 2025–2026 alone.
Options on the same index publish their own probability every Monday morning: the price of a tight put spread at a strike is, to a good approximation, the market's implied odds that the week settles below it. If the market's implied odds are persistently higher than the physical odds at a specific level, selling that level collects the difference.
So we asked the market directly. For each of the 327 weeks from May 2020 to August 2026, we took SPXW quotes at 10:00 ET on the week's first trading day, at strikes sitting on each weekly ATR level — five levels up, five levels down — and compared implied probability against what actually happened by Friday settlement.
Share of weeks whose Friday close finished beyond the level. Three independent windows — the stability is the point:
SPY = 1,366 weeks (2000→2026). SPX = 328 completed weeks (2020-05→2026-08). Levels built from prior-week close ± fib × prior-week ATR(14).
Implied probability = mid price of the 5-point vertical at the level strike ÷ 5, at Monday 10:00 ET. Realized = share of weeks that settled beyond that same strike interval.
The −1.0 ATR put cell, year by year. The gap column is implied minus realized.
A put option is an insurance contract on the market. The buyer pays a premium up front. If the S&P 500 ends the week below an agreed price (the strike), the seller pays the buyer the difference. If it ends above the strike, the seller keeps the premium and owes nothing.
We are the seller. We sell one put at the weekly −1 ATR level — the line the market closes below in only ~5% of weeks. Selling it alone would work most weeks, but one terrible week could do real damage. So at the same moment we buy a cheaper put 50 points lower. That second put is our own insurance: below its strike, whatever we owe on the put we sold gets paid back to us by the put we own. The loss stops there, guaranteed.
The two-legged position is called a spread, because you're no longer exposed to the whole market — only to the 50-point gap between the two strikes. The specific kind traded here is a bull put spread, also known as a put credit spread or short put vertical: "put" because both legs are puts, "credit" because cash lands in your account on day one, "bull" because it wins whenever the market holds up, and "vertical" because the legs differ only in strike.
That's the entire deal: a high-probability small win against a rare, capped loss. The whole study is about whether the small win is priced generously enough to carry the rare losses. It is — because the market persistently overpays for insurance at this specific level.
round5(prior weekly close − 1.0 × prior-week ATR14) — the weekly −1 ATR level, rounded to the nearest 5.Fills in the backtest are at the quote midpoint on both legs, minus $2.64 commission per spread. Average credit collected: 2.98 points ($298) against a max risk of roughly $4,700–5,000 per lot. Median displayed size at the short strike was 89 contracts at entry time. For the floor: the worst-case fill assumption (sell the short at the bid, buy the wing at the ask, never split a spread) still produces PF 2.09 and +$40,170 — real fills land between the two.
All rows: same entry day, same fills, same commissions. Only the strike rule differs.
Why the control matters. A fixed-distance rule (strike always 2.74% below the prior close — the median ATR distance) collects more credit but hands it back in volatility regimes: 2022 was a losing year for the control and its max drawdown is nearly 3× larger. The ATR strike backs away from price exactly when vol expands. One honest caveat from the audit: the fixed control sits at a nearer effective delta (implied ~9.8% vs ~7.7%), so this comparison shows ATR beats the naive rule — it does not isolate ATR magic from "sell a ~7% implied-probability put weekly." A delta-matched approximation still favored ATR (PF 1.54 vs 2.09).
Calendar-year P&L per 1-lot, primary strategy vs control.
The whole sweep, same rule, only the wing distance changes:
15 losses in 326 trades. Four settled through the long wing and paid the full 50 points — those four are nearly half of all losing dollars. Fourteen of the fifteen were grind-down weeks (intraday drift through the strike), not overnight gaps; the one true gap was the April 2025 tariff week.
Week-max VIX separates losers from winners far better than entry VIX (median 29.0 in losing weeks vs 21.5 in winning weeks). Entry-time VIX is not a usable gate — losers and winners start from similar levels. A trend gate doesn't help either: splitting entries by SPX above/below its daily 21 EMA gives PF 2.14 above vs 2.04 below. The filter halves the trade count and the profit and improves nothing.
The full study was adversarially audited by an independent model (GPT-5.6) with shell access to the same data. The audit ran on the worst-case-fill version of the backtest (sell at bid, buy at ask — PF 2.09, +$40,170) and reproduced it to the penny; this page reports midpoint fills, which shift the level of the numbers but none of the conclusions. Confirmed: point-in-time level construction (zero leakage found), settlement values (306 of 308 checkable expiries match Cboe's official SPXW settlement archive exactly), holiday-week handling, and five hand-reconciled trades including both largest losers.
It also corrected the write-up in four places, all reflected on this page:
Statistical strength, stated plainly: realized loss frequency 4.67% against implied 7.7% is a one-sided binomial p of 0.021 over 321 trades; the paired weekly gap bootstraps to a [+0.7, +5.3] point interval. The per-year positive streak is descriptively true but thin — one extra losing week would have flipped 2021 or 2023.