← sources

My LSTM is fine at 1 day ahead and useless at 1 month. Is that just how price series work?

source: reddit (reddit:deeplearning)published 2026-11-18 13:20ingested 2026-11-19 02:36

retained text

student project, forecasting crypto price ranges (btc, eth, sol) with an LSTM that outputs quantiles.

\- 1 day ahead: the 80% band covers about 76%, decent

\- 1 week: 68%

\- 1 month: 51%, basically a coin toss with extra steps

i tried more layers, more history (2 years instead of 6 months), and adding volume and funding rates. nothing moved the 1 month number much.

my guess is the 1 month range depends on stuff that just isnt in the price history (news, macro). but i might also be doing something dumb with how i build the training windows.

is it normal for DL models to fall apart like this on longer horizons, or should i be looking at a different architecture?

document fetched 2026-11-19 00:24 · canonical https://reddit.com/r/deeplearning/comments/1xzf17y/my_lstm_is_fine_at_1_day_ahead_and_useless_at_1