I Deployed a Live AI Trading Bot and It Called the BTC Crash in Real Time

There's a specific feeling when something you built actually works in production. Not in a test. Not in a backtest. In the market, with real money, in real time.

Today was that day.

What We Built

Over the past few weeks, I've been building two systems that were designed to run together:

1. PolyEdge — an AI trading bot that evaluates Polymarket binary options markets using a 9-factor scoring system (InDecision framework as the backbone, fed by live Coinbase price data, MACD, volume, Chainlink oracles, and window timing)

2. Mission Control — a real-time dashboard that shows me everything: wallet balance, open positions, trade history, P&L, bot health status, and the score that drove each decision

The trading bot runs on a Mac Mini 24/7. Mission Control is served through a Docker container behind a Cloudflare tunnel. I can pull it up from anywhere and see exactly what the bot is doing and why.

Today I pulled it up and watched it catch a move.

The Setup

BTC was already in a confirmed downtrend. Overnight it had dropped from approximately $69K to the low $65K range — about a −6.5% macro move. The 4-hour MACD reading when I checked the chart: −325.54 on the MACD line, −589.94 on the signal. The histogram expanding. No reversal signal. Just sustained, directional selling.

This isn't noise. This is structure. When you see MACD this extended on the 4h with no divergence and no volume floor, the trade isn't "figure out where to buy." The trade is "don't fight it."

The First Win — 10:08 AM

The bot's first trade of the day: BTC DOWN at 49¢ on a 15-minute Polymarket binary window.

DOWN tokens at 49 cents. That means the market was pricing BTC's 15-minute move as nearly a coin flip. But it wasn't a coin flip — BTC was in a confirmed downtrend with active selling volume at 2× normal rate and every technical indicator pointing the same direction. The market gave us near-even odds on an aligned thesis.

Result: WIN. Net P&L: +$4.38.

The Bounce Trap — 11:21 AM

One hour later, the bot placed a second trade. This one I want to talk about at length because it's more instructive than the win.

BTC UP at $0.94. A brief 15-minute recovery candle had printed in the middle of the crash. The momentum signal scored enough to pass the threshold. The market priced UP tokens at 94 cents — implying 94% probability of an UP close on that window.

The bot bought the bounce.

Here's the math on why that's wrong before we know the outcome:

You're paying $0.94 to win $0.06. That's a gross return of 6.4%. After Polymarket's 10% fee (applied to the bet size), your fee is $0.60. You made $0.60 gross on the win. The fee ate the entire profit. Net P&L: −$0.22.

The direction was technically correct. The market closed UP. But the odds were so crowded — so late in the momentum move — that fees consumed the edge entirely.

The Mission Control dashboard now shows this trade as BREAK EVEN — an amber badge I added specifically for wins that lose money after fees. Because "WIN" felt dishonest. The direction was right. The trade was wrong.

The Rule We Added in Real Time

After the retro on that trade, I updated both strategy files before the bot's next evaluation cycle:

if bet_price >= 0.85:
    skip_reason = "Extreme market odds — bounce trap risk, near-zero EV"

Hard rule. No score override. If the market is pricing one outcome at 85% or higher, skip it regardless of what the signal says. Someone on the other side of a 94¢ bet knows something you don't — or they're about to find out you were wrong together.

The rule went live in under 20 minutes. The bot restarted. The next evaluation cycle was already running under the new regime.

The Precision Call — ~12:15 PM

With the extreme-odds rule live, the bot entered the early afternoon. BTC had continued to slide. The 15m MACD was −149.73 / −313.80 with the histogram at −163.87 and expanding. Price had dropped to $64,529 — down 4.57% on the day.

The bot placed a fresh BTC DOWN position. This time at fair odds. DOWN tokens at 55¢.

The signal breakdown:

• Momentum: 30/35 (strong directional delta)
• Technical analysis: 14/20 (MACD confirmation, no counter-signal)
• Volume: 2× normal (active selling, not drift)
• Extreme odds: NOT flagged — fair market, no bounce trap pattern

This is exactly what the system was designed for. Not a crowded 94¢ signal. A 55¢ signal with actual edge. Multi-timeframe MACD alignment, volume confirmation, direction aligned with the macro structure. Clean.

When I pulled up Mission Control at 12:35 PM, I saw this:

• USDC.e Balance: $47.98
• Total P&L: +$4.16
• Win Rate: 100%
• Open Positions: 1 (the BTC DOWN from 12:15 PM)

BTC was at $64,455 and still falling.

What This Means

I'm not posting this to tell you our bot is perfect. The bounce trap trade proves it isn't. I'm posting this because the feedback loop matters more than any individual trade.

We deployed. We watched it trade. We identified a flaw (extreme odds handling). We patched it in real time. The next qualifying trade was cleaner, more precise, and aligned with what the framework is actually supposed to do.

That's the advantage of building your own system. You don't wait for a vendor to patch the blind spot. You see it, you fix it, you move on.

The InDecision framework now has a live trade record to defend. The bounce trap is in the permanent loss ledger. The DOWN precision call is in the win column. Both are real. Both are instructive. Both are in the database.

The bot is running. The dashboard is live. The edge is measurable.

More to come.