60k live trades taught me how not to build a trading bot. rebuilding it in public. market microstructure · execution · node/ts · go
Joined November 2023
- Tweets62
- Following41
- Followers5
- Likes47
While building the first version of my bot, I spent a lot of time reading Reddit, forums and algo trading communities.
I noticed a pattern.
Beginners obsess over strategy.
Win rate.
Entries.
Indicators.
Backtests.
Almost nobody talks about where the bot actually runs.
For short-term trading, putting your server in the wrong place can add enough latency to completely change your fills - and kill an edge that looked great on paper.
A good strategy with bad execution is a different strategy.
Before coding a couple of new bots, I decided to trade their strategies manually on OKX first.
You know, just to make sure they actually make sense.
What could possibly go wrong?
Ran my bot for 6 weeks.
60k trades.
Closed P&L: green.
I was pretty happy with that number.
Then I marked every open position to market.
⬇️
The picture changed completely.
The problem wasn't necessarily the strategy.
It was the measurement.
Winners closed automatically.
Losers could stay open.
So my closed-trade history wasn't a sample of my strategy's outcomes.
It was a sample of the outcomes my exit logic had chosen to finalize.
I wasn't measuring my strategy.
I was measuring my exit logic.
High win rate.
Smooth closed-P&L curve.
Thousands of small green trades.
Even a random entry signal can look surprisingly good when winners are realized and losers are allowed to sit.
The fix isn't clever.
Mark-to-market has to be the primary P&L metric.
And every position needs to resolve within the horizon the signal actually lives on.
If your alpha decays in seconds, a position that's still open six hours later isn't your alpha anymore.
I'm rebuilding the measurement layer before touching the signal.
Next I'll post what the honest equity curve looks like.
This made me rethink how I look at order flow.
It's easy to treat volume and trades as inputs.
But if part of that flow is actually feedback - one event increasing the probability of the next - then you're not just measuring activity.
You're measuring how close the market is to a cascade.
5 years holding.
5 years waiting for “the next bull run.”
But who said there had to be one?
Crypto isn't a baby market anymore.
The easy money gets arbitraged away.
The obvious cycles stop being obvious.
The market never promised you another bull run.
92% win rate.
Sounds like a great trading bot.
Until you realize most of the losing positions were never closed.
My biggest bug wasn't in the code.
It was in the metric I trusted.
🤖 Made with AI
When I started building a trading bot, I obsessed over indicators, patterns and signals.
I barely thought about funding. Even less about fees.
Turns out those boring costs can kill a great strategy.
And when you're trading short-term market microstructure, a few bps can be the entire edge.
I thought my trading bot had an exit strategy.
It didn't.
It had a take-profit and a liquidation condition.
After ~61k live trades, that difference left most of the portfolio stuck in positions.
And somehow, the bot still showed a 92% win rate.
The rebuild starts with exits, state and testing.
deterministic event replay
exchange/local state reconciliation
tests around every money-moving path
explicit forced-exit logic
different passive/aggressive exit models