Deflow After 11 Trading Days: Right on Small Caps, Wrong on Large Caps

At a glance

Most of the loss is still open. The closed trades are profitable, but the six spreads still on the book are marked $12,779 below where they were sold. Whether this turns into a realized loss depends on what the market does before those spreads are closed or expire.

What Deflow is designed to do

Deflow was built by Ifeanyichukwu Onwo as a finalist entry in the lablab.ai × Alpaca AI Trading Agents Hackathon (source on GitHub). It runs like a small options desk, cycling every five minutes during market hours across SPY, QQQ, IWM, NVDA, AAPL, MSFT, AMD and TSLA:

• An analyst measures the variance risk premium: 30-day at-the-money implied volatility minus a forecast of realized volatility. It combines that with IV rank and the 20/50-day trend to classify the regime.
• When options look expensive (at least 2 volatility points of premium and IV rank of 40% or more), the desk sells premium: a bull put spread if the trend is up, a bear call spread if it is down, or an iron condor if it is flat.
• When options look cheap (premium of −1 point or less and IV rank of 55% or less), it buys a bull call or bear put debit spread. Otherwise it stands aside.
• A structurer prices eight candidate spreads, and a language model picks one or abstains. The model cannot change prices, strikes or size. The original entry used Qwen2.5-72B; we run the strategy with Claude, through our Claude subscription.
• An auditor re-derives the Greeks and runs a 1,000-path jump-diffusion Monte Carlo simulation. It rejects any trade with zero or negative expected value.
• A deterministic risk gate with twelve breakers has the final say. Among other limits, it caps maximum loss at 2% of equity per trade, 3% per symbol and 6% for the whole book.

Exits are rule-based. A spread is stopped out at 50% of its maximum loss, profits are taken at 75% of maximum profit (sliding to 40% as expiry nears), and everything is closed with three or fewer days to expiry.

Results by underlying

Paper account, October 2–16, 2026. Realized P&L is from Alpaca fill activity; open P&L is each position at its current mark.

The desk only traded the three index ETFs in this period. None of the five single stocks produced a trade.

What happened

The open book tells the story. On IWM the desk sold bear call spreads, a bet that small caps would not rally. On QQQ and SPY it sold bull put spreads, a bet that large caps would not fall.

Credit and current value are per share for the spread as a whole. Max loss is the spread width less the credit received, times contracts × 100.

The QQQ spread is the problem. The short 725 put was sold for $7.02 and is now marked at $22.17. QQQ closed Friday at $709.80, below both strikes, so if it stays there the spread will be worth its full $15 width at expiry and the maximum loss will be realized. The spread sold for $2.48 is now worth $9.10. SPY closed at $763.25, still above its short strikes of 756 and 752, so its two spreads are down much less.

Taken together, the book was positioned for large caps to hold up and small caps to stay capped. From the October 9 close to October 16, QQQ fell 5.5%, IWM 3.1% and SPY 2.0%. Small caps fell, so the IWM bear call spreads worked. The bullish large-cap spreads did not, and QQQ, the most volatile of the three, fell the furthest.

The losing trades are not independent of each other. SPY and QQQ usually move together. Two bullish put spreads on them are close to one larger bet on the same outcome, and the per-symbol limit does not treat them that way. We look at this in more detail in Are SPY and QQQ Credit Spreads One Bet or Two?

Where the live book differs from the design

• Size per trade. The design caps each trade's maximum loss at 2% of equity. Measured against the report's $110,189 capital base, that is about $2,200. The QQQ spread can lose up to $20,032 (18%), and the larger SPY spread up to $19,720 (18%).
• Book total. The design caps the whole book at 6%. The six open spreads have a combined maximum loss of $60,030, about 54% of the capital base.
• Stop-loss. The design closes a spread at 50% of its maximum loss. At the Friday close the QQQ spread was down $6.62 per share against a maximum of $12.52, about 53%, and it was still open. A stop checked every five minutes could have been crossed in the final minutes of the session. Next week's fills will show whether it triggers at the open.

These are differences between the deployed configuration and the published description. They are not necessarily flaws in the strategy's logic. A desk that sizes as designed would have lost roughly a tenth as much on the same trades.

Limitations

• Eleven trading days is far too short. Premium-selling strategies typically show many small gains and occasional large losses. Two weeks can land on either side of that distribution and says little about the long run. Annualized figures over this window are not meaningful.
• Paper fills. These are paper-trading results. Live multi-leg fills can be worse, especially on wider spreads.
• Daily figures are cash-flow based. The performance report records realized P&L when fills occur, so a day's figure can reflect premium collected on opening a trade rather than a gain. We rely on the totals and the open-position marks, not the day-by-day path.
• The regime model is still warming up. The strategy ranks implied volatility against observed implied-volatility history only after 20 sessions. Until then it uses historical volatility as a proxy, so its "options are expensive" signal has been running on its fallback measure.
• No backtest. There is no historical backtest for Deflow on the platform yet, so there is nothing to compare these two weeks against.
• The return base. The −7.58% figure uses the report's $110,189 capital base, which is derived from the positions traded rather than the account's equity.

What we will watch next week

• Whether the QQQ spread is closed by its stop-loss, and at what price. It must be closed by October 27 under the three-days-to-expiry rule.
• Whether new trades are sized within the published 2% limit.
• Whether the desk keeps selling bullish premium on SPY and QQQ while they trade near or below the short strikes.
• How the 35-lot IWM bear call spread, now the largest position by contract count, holds up if small caps rebound.