It's been quiet around here.
There's a reason for that.
Back in March, we wired 10 independent data sources into a fusion engine, pointed it at a $100,000 Alpaca paper account, and told it: go. No human override. No morning checklist. Every 30 minutes, check your feeds, evaluate your signals, and if the math says go — go.
Then we did the hardest thing: we let it run. No tweaks. No "just this one adjustment." We wanted clean data. Two months of it. Enough to look at the numbers and say something real about whether any of this works.
Today is Day 61.
Let's look at the numbers.
Yes — the bot made money. That's genuinely not nothing. A lot of trading systems blow up in their first two months. We didn't. The account is alive, uninjured, and green.
But if you'd taken the same $100,000 on March 28, put every dollar into SPY, closed your laptop, and spent the next 61 days doing literally anything else — you'd have made $2,864. That's 4.7 times what Tauntaun produced.
If you'd put it into QQQ? $3,373. Five and a half times.
We ran 10 data sources. We parsed 8 RSS feeds. We polled FRED economic indicators, tracked geopolitical risk indexes, watched prediction market whales, scanned SEC insider filings — and the entire elaborate contraption got beaten by one ticker symbol and a nap.
This is the part where we're supposed to be honest. And the honest truth is: this is fascinating.
The Win Rate Problem
Of the 39 trades that closed with full P&L data, 19 were winners. That's 27%. If you're keeping score at home, that means 73% of trades lost money.
How did the account stay green? Because the winners were bigger than the losers. The system's edge — if you can call it that — was asymmetry.
The best trades:
- URA +17.1%
- IWM +11.9%
- XAR +6.9%
- HACK +8.3%
- USO +10.0%
And the worst:
- VEA −14.7%
- VXX −13.3%
- XLK (short) −13.2%
- VXX −9.2%
- UNG −8.5%
The pattern is hard to miss. The winners came from sector rotations and small-cap momentum (nuclear energy, aerospace, cybersecurity, small caps ripping). The losers came from volatility bets and short-selling in an uptrend. VXX was a serial offender — the system kept trying to trade fear and kept getting burned.
And the shorts? Four SPY shorts, three XLK shorts, multiple VEA shorts. Nearly all of them lost money. In a market that's been grinding up (SPY +2.86% over the period), the bearish signals were fighting the tide. The regime filter was supposed to catch this — suppress new shorts when the trend is bullish — but clearly it wasn't aggressive enough.
500 Decisions. 39 Trades. A Lot of "No."
The system processed 500 signal events in 61 days. It said yes 54 times and no the other 446. That selectivity is actually the best thing it does — the gatekeeping kept us alive. The problem isn't the gate. It's what got through.
Every decision is logged with its signal chain — which sources fired, at what confidence, with what cross-theme bonus. You can trace every trade back to a specific set of inputs. When URA ripped to +17%, that was an RSS cross-theme spike (nuclear energy × military conflict) combined with a GPR crisis reading. The system was right. When VEA short collapsed to −14.7%, that was a reactive trade on a weakening-dollar thesis that the market simply didn't agree with. The system was wrong.
The difference between those two trades is what we're here to figure out.
What Actually Happened Under the Hood
The trailing stop system did its job. Almost too well. Look at the exit reasons on the closed trades and you'll see the same story over and over: "trailing stop hit." The ratchet mechanism worked — URA ran up, locked gains at +8%, tightened the stop, and exited at +17.1%. Beautiful.
But it also dumped positions early. USO got stopped out at −5.7%, −4.4%, −1.6%, −1.2% — four separate times. XLE, four stops. VXX, four stops. The system was excellent at getting out but terrible at staying in the things that would have recovered.
We have 16 open positions right now. XAR is up +11.6% unrealized. ITA +8.9%. QQQ +30.1%. Some of the longs are working beautifully. But we also have −$66 on a SPY short and −$134 on a VEA short — positions that should never have been opened in a bull market.
Signal selectivity — 446 rejections out of 500. The trailing stop ratchet on winners. Cross-theme scoring (oil × military conflict, nuclear × geopolitics). The 10-source fusion design itself — the architecture is sound. Not losing money.
Short-selling in an uptrend — the regime filter needs teeth. VXX / volatility products — stop trading these. Premature stop-outs on sector positions that later recovered. The win rate (27%) — we're taking too many low-conviction trades. Over-trading: $33,838 invested across 16 positions is 34% of the book. For a system still calibrating, that's too much.
What This Means (And Doesn't Mean)
This is not a failure. It is a v1 delivering exactly what v1s deliver: it survived, it generated real data, and it revealed exactly where the structure breaks.
If we'd made $5,000, the takeaway would be "ship it to production." That would have been the wrong lesson. A 27% win rate that lucked into a couple big trades is not a system — it's a lottery ticket with extra steps.
If we'd lost $5,000, the takeaway would be "shut it down." Also wrong. The signal architecture did find real edges. URA at +17%, IWM at +12%, HACK at +8% — those weren't accidents. The fusion engine spotted those themes before they moved. The problem is that it also spotted 52 things that didn't work.
The truth is in the middle: +$609, on a system that processes 10 independent data sources every 30 minutes and trades completely autonomously, is a real result. It's just not a good enough result. And we now know exactly where to fix it.
Where We Go From Here
1. Kill the shorts. In a market grinding upward, the short book is just a slow bleed. The regime filter needs to be binary: bull regime = zero new shorts, period. No exceptions, no safe-haven carve-outs. If the thesis is bearish, sit in cash.
2. Stop trading VXX. Volatility ETPs are not investments. They are instruments for expressing a specific view on implied vol, and our system doesn't have a vol model. It's just pattern-matching "fear" from RSS feeds into a decaying product. Delete the row from the config.
3. Raise the confidence floor. The current minimum is 30%. At 30%, you're taking trades on a coin flip plus a light breeze. Push it to 50% and see what happens to the win rate. A system that takes 10 trades and wins 6 is worth more than one that takes 39 and wins 10.
4. Lengthen the trailing stops. We're running 5% ATR stops. At least three sector positions got stopped out, only to recover within days. The ratchet mechanism on winners is excellent — keep it. But the initial stop width on new entries needs room to breathe. Try 8%.
5. Position size discipline. 34% of the book invested is too much for a system with a 27% win rate. Cap it at 20% total exposure while we're recalibrating. The cash isn't doing nothing — it's keeping us alive.
The Real Lesson
We built a system that reads the news, watches the Fed, tracks geopolitical risk, monitors prediction markets, scans SEC filings, and fuses it all into trading decisions — and it survived two months without blowing up.
That's the good news.
The bad news is that buying SPY and going to the beach did 5x better.
But here's the thing: SPY doesn't teach you anything. SPY doesn't generate 500 labeled decisions you can dissect. SPY doesn't tell you which signal sources have predictive power and which are just noise. SPY doesn't build toward a system that, on v7 or v12, might actually see things before the market does.
We now have 61 days of clean, un-tweaked data on every signal, every trade, every exit. That's the real asset. The $609 is just the receipt.
The work on v2 starts now.
Still paper trading. Still learning in public. Still not dead. 🦎