Variance for Option Sellers: Surviving Losing Streaks

By Charlie Nguyen · Published Nov 5, 2026 · Model data Jan 3, 2011 to Oct 23, 2026 · 12 min read

Variance is why a high-win-rate credit spread seller still has bad weeks. An options trading losing streak is not a signal that the edge vanished; it is the price of sampling a 90%+ process. Here is the closed-form math for the longest losing run, what a 15-year SPX put credit spread model actually did, and how to size so a streak does not end the account.

The math of a losing streak

Treat each trade as a win with probability p or a loss with probability q = 1 − p, independent of the others. After n trades, let L be the longest run of losses in a row. Then:

• E[L] = Σ_k≥1 P(L ≥ k) — the expected length of the longest losing streak.
• P(L ≥ k) comes from a short Markov chain: states track the current run of losses; state k means you have already hit k losses in a row.

That is the same geometric-run math behind coin-flip streaks. It does not need market data. It only needs a win rate and a trade count. Markets are not perfectly independent, so later I also measure the streaks the SPX model actually produced.

What a 90%+ win rate still allows

Fix n = 250 trades (about a year of almost-daily 7 DTE entries) and vary the win rate:

What it shows. At a 96% win rate, you should expect a longest losing streak of about 1.33 trades. Roughly 32% of the time you will see at least two losses in a row; three in a row is about 1.5%. Drop the win rate to 90% and three-in-a-row jumps to about 20%. A Monte Carlo of 40,000 iid paths matches the closed form to two decimals.

More trades, longer expected streaks, even at the same win rate:

At 96% and 1,000 trades, E[L] rises to about 1.85 and the chance of two in a row is about 79%. Volume does not erase variance; it makes the rare streak almost inevitable.

How I tested it on SPX

Every trading day from Jan 3, 2011 to Oct 23, 2026, the model opened two SPX put credit spreads at the close:

• 7 DTE, 9 delta, $50 wide (5 trading days). Median model credit $1.73, short strike a median 3.1% below SPX.
• 45 DTE, 10 delta, $20 wide (31 trading days). Median model credit $0.93, short strike a median 8.5% below SPX.

Each spread was closed by one of four exit rules: hold to expiration, a 50% take-profit (the 50% rule tested), a 2× credit stop (stop losses on put credit spreads), or both. I then measured losing streaks two ways: on the overlapping daily-entry series, and on a one-at-a-time series that only re-enters after the prior trade exits.

Model backtest, not real fills. Option prices are Black-Scholes, with implied volatility from Cboe’s VIX9D (0.77 × VIX9D) for 7 DTE and VIX (0.73 × VIX) for 45 DTE, plus a put skew fitted to a real SPX put chain. Risk-free rate from the 13-week T-bill, dividend yield 1.8%. Fills at the model mid, checked at the daily close only. Fees $2.60 per spread to open and again to close early. P&L is per one spread ($100 multiplier). Results are hypothetical.

Results: 7 DTE spreads

*Months of EV wiped = absolute value of the worst one-at-a-time losing streak, divided by (P&L per day × 21 trading days). Model dollars per one spread.

What it shows. Holding with no stop won 96.0% of one-at-a-time trades (794 trades). The closed-form expected max streak at that win rate is 1.76; the model observed 2. That worst streak lost −$3,051 — about 11.9 months of the model’s EV per day. A 50% take-profit looked similar on streak length (2) with a slightly higher win rate. A 2× stop dropped the win rate to 80.7% and raised the expected max streak to 3.85 (observed 3).

Results: 45 DTE spreads

*Months of EV wiped = absolute value of the worst one-at-a-time losing streak, divided by (P&L per day × 21 trading days). Model dollars per one spread.

What it shows. One-at-a-time sample sizes are smaller (128 holds), so a single max loss dominates the “worst streak” column. Holding won 99.2% with an observed max streak of 1 (theory 0.64). The 2× stop again lengthened count-streaks (observed 4, theory 4.02) while cutting EV per day to $0.83 from $1.22.

Overlapping trades vs one at a time

Independence is the closed-form assumption. Real entries are not independent when you open a new spread every day:

• 7 DTE hold, overlapping: 3,970 daily entries, observed max losing streak 6.
• 7 DTE hold, one at a time: 794 trades, observed max losing streak 2.

A crash week can mark several overlapping spreads as losers at once. That is why the overlapping series looks scarier than the Bernoulli math. If you run more than one spread at a time, plan for clustered losses, not just iid streaks (more on bankroll and risk of ruin in bankroll management for +EV options).

Do stops or take-profits change streaks?

Two companion articles already measured EV. Here is the streak angle only:

• 50% take-profit (when to take profits): slightly higher win rate on short-dated spreads, so expected count-streaks get a little shorter. It does not remove the crash cluster.
• 2× credit stop (what the stop-loss data says): lower win rate, longer expected count-streaks, smaller dollars per loss. You trade more frequent small losers for fewer full max losses.
• Both together: the tightest dollar package in the stop-loss test, and the longest expected count-streaks of the four rules on 7 DTE (3.99).

Pick the exit for the dollar distribution you want. Do not expect a stop to make losing streaks “go away” in the win/loss tape.

How a streak wipes months of EV

On the 7 DTE hold rule, average P&L per day in the trade was $12.25. The worst one-at-a-time losing streak lost −$3,051. Divided by 21 trading days of that EV, that is 11.9 months of grind erased by 2 losers in a row.

That is the variance trap: the average trade is small and positive; the rare loser is many times larger. A high win rate does not make the account safe unless size respects the loser, not the winner.

Sizing so a streak does not end the account

Work backwards from a streak you are willing to survive. Example (illustrations only, not a recommendation and not anyone’s real account):

1. Choose a max risk per spread as a fixed slice of the account (1%, 2%, …).
2. Ask what three full max losses in a row would do (at 96% win rate, P(≥ 3) over 250 trades is about 1.5%, small but not zero).
3. If that drawdown is unacceptable, cut size until it is.

Here the model’s actual worst 7 DTE one-at-a-time streak (two losses, not both full max) is scaled to example accounts. The dollars at risk are the chosen fraction of the example account; the streak P&L scales each model loss by (chosen risk ÷ that trade’s model max risk):

EXAMPLE accounts and risk fractions only. A streak of full max losses at the same fractions would be larger (2 × risk fraction for two max losses; 3 × for three). Gaps can deliver that. Size for the streak you refuse to let end the account.

Example: a live spread and its odds

Here is the SPX Desk’s 7DTE model spread from the Nov 4, 2026 close. POP is the chance of any profit at expiration; a 90%+ POP is exactly the kind of win rate that still produces losing streaks over a year of trades.

EXAMPLE, model figures. Cboe delayed SPXW quotes at the close, credit at the mids, SPX 7,688.62. POP = chance of any profit at expiration; POT = chance SPX touches the short strike first; P50 = chance the spread can be bought back for half the credit. Lognormal model at the short strike’s IV 16.6%, rate 4.03%, P50 from 40,000 simulated paths. See POP vs POT vs P50 and IV rank explained.

My variance rules

1. Assume the streak will show up. At 96% and a few hundred trades, two losses in a row are common; three are rare but real.
2. Size for three max losses, not for the average winner. If three full risks in a row would knock you out, the size is too big.
3. Count overlapping risk separately. Correlated positions can lose together even when the one-at-a-time streak math looks mild.
4. Use stops for dollars, not for streak vanity. A 2× stop lengthens count-streaks while shrinking each loss (data).
5. Keep short-dated take-profits for EV per day (50% rule), knowing they barely change streak length.
6. Pick strikes outside the expected move with Open Lines (why), and sell more when IV rank is high (IV rank).
7. Never blow up. The edge only pays traders who are still in the game after the streak.

Sources and method. Closed-form longest-run probabilities from a Markov chain on the current loss run (standard Bernoulli run theory; Monte Carlo check with 40,000 iid paths, seed 20261105). SPX, VIX, VIX9D and 13-week T-bill daily closes from Yahoo Finance (Jan 3, 2011 to Oct 23, 2026); VIX and VIX9D calculated by Cboe. One-at-a-time series re-enters at the next close after each exit. Win = P&L above zero after fees. Model figures, not historical option quotes; results are hypothetical. Example: Cboe delayed SPXW quotes (Nov 4, 2026 close); SPX terms from Cboe SPX specifications. Code: raw/variance/backtest.py.

FAQ