TikTok forecasting
In progress · First results targeted for Dec 2026
Forecasting a video's views with conformal intervals that say how sure the model is.
Problem
A creator deciding what to post wants to know how a video is likely to do and how sure that guess is. A single predicted view count hides the uncertainty. This project forecasts views with conformal prediction intervals, which cover the true value at a chosen rate on held-out data.
Done so far
- Forecasting code and a conformal-interval benchmark for the public SMP-Video dataset.
- Code to load the public TikTok-10M dataset, split it into train and test sets without leaks, and rank models on a shared leaderboard.
Approach
Pretrain on public data, then adapt
Models train first on public TikTok data (SMP-Video and TikTok-10M), then adapt to a single creator's account, where there are far fewer videos to learn from.
Conformal intervals
Conformal prediction wraps any point forecast in an interval, calibrated on held-out videos so it covers the true view count at a rate you choose, without assumptions about the model.
How it fits together
- Public TikTok datasets
- Leak-free splits
- Forecasting model
- Conformal calibration
- Forecast with an interval
Status
- No results yet. Coverage and interval width will be reported once the models are trained.
What's next
- Download TikTok-10M and train the first models on it.
- Adapt the pretrained model to one creator's account.
- Report coverage and interval width on held-out videos.