Repository navigation
Same number of price bars, wrong dates: an offline reliability example #1442
Unanswered
arian-gogani
asked this question in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
i'm building Nobulex, an open-source reliability suite for financial-data tools. here's a small failure case that seems relevant before data reaches a training or backtesting pipeline.
a request returns five price bars. the reference also has five. a count-only check agrees, even when those bars cover entirely different dates.
the existing offline example uses fictional data and shows three outcomes from the suite's truncation classifier:
that last result means this probe couldn't establish a clean comparison. it isn't a pass. even the matching-session PASS doesn't establish that the prices themselves are correct.
The example and its limits are here. It runs offline after cloning, with Python 3.11 or newer and no credentials.
this isn't a FinRL integration or a finding about FinRL. i'm interested in the data boundary before the model: how do you check that the returned sessions match the requested window, especially when the response looks successful?
if you have a fictional or shareable counterexample where this check gives the wrong outcome, that would be useful feedback. please keep private datasets and credentials out of replies.
All reactions