# I open-sourced 2,000 more dead crypto bots. My best ideas did no better than my random ones. — retained record

> Retained reddit record from reddit:algotrading.

Canonical: https://highsignal.app/data/records/91d9c6a3de59718b

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:algotrading)  published 2026-11-11 14:24 ingested 2026-11-12 02:43

[original source ↗](https://www.reddit.com/r/algotrading/comments/1xtj1lw/i_opensourced_2000_more_dead_crypto_bots_my_best/)

## retained text



```
Since the last time, I made lots of changes, many of you told me my judge wasn't that great nor trustable! So before dropping more bots, I fixed what you pointed at:

* Placebo screening, 500 random-entry bots through the exact same screen. 0 of 500 passed and I ran it again, 0 again.
* Real costs: 4bps taker fee plus impact/spread on every side.
* I realized you guys just had to take my word for all the tests and results so I open-sourced the validation stages and also added a deeper validation stage (6+ years) that I apply to all the bots after this 7 stage validation. Go judge my judge!

Then I ran this new massive batch for days. New ideas, branches of previously better performing bot ideas and wildcards.

[github.com/CacheCarti/Crypto-Strategies](http://github.com/CacheCarti/Crypto-Strategies)

What happened:

* \~1,980 bots went through the screen. 6 survived (0.3%). The other 1,970 are in the repo with a report card each, showing exactly which test killed them.
* I split the batch into three buckets: exploit (variations on stuff that already worked), explore (new ideas), and wildcard (deliberately random/weird).
* How they all died🥀: 1,197 in-sample (60%), 668 out-of-sample, 98 on robustness (randomized starts 48, perturbation 40, walk-forward 10), 29 ay holdout, 2 failed code check. 6 lived.
* My "best" bot made 8% in 6 months "so that's 16% a year, with 2x leverage that's 32%!!" NOOO don't fall for this, it WILL die the second it hits real money. That is the lesson of my entire experiment...specious results!
* Exploit passed 4 of \~1,200. Explore 1 of 500. Wildcard 1 of 300. So my carefully aimed ideas did about as good as the random ones.
* Got a new VM for the validation pipeline for public use not for me. You can validate any of your bots and get a certificate with code hashed.

Things you should know before you trust any of it:

* The card screen only uses \~6 months of data. It's the cheap first filter, not the exam. The 6 survivors go through testing on years of data.
* These bots were judged on an earlier version of the judge. I'm getting the validator externally audited and shipping improvements to it constantly. Please read the validation code and please try and see it's gaps!

If you think a test is missing, or if the judge has gaps please tell me. Every suggestion so far has killed more of my bots and honestly made it much better imo! That's still the process!
```



document fetched 2026-11-12 00:46 · canonical https://reddit.com/r/algotrading/comments/1xtj1lw/i_opensourced_2000_more_dead_crypto_bots_my_best
