Calibration of YouTube View Forecast Model

Subject: Quote request. Probabilistic calibration of a forecasting model for expected-value calculation

Hello,

I am looking for a statistician or a small specialised firm for a
fixed-price engagement. I would like a single quote for the whole work.

GOAL
Correctly calculate the expected value of bets on prediction markets.
Expected value comes from comparing the estimated probability of each
outcome range with its market price: if that probability is poorly
calibrated, expected value is wrong even when the central estimate is
accurate. Calibration is therefore the core of the engagement.
I am not asking for a certain prediction of the outcome, but for
probabilities calibrated as well as the available data allows, with their
uncertainty measured and stated.

CONTEXT
The system forecasts the view counts of videos from a large YouTube
channel from 24 hours to one week after publication. Markets pay out by
view-count range, with day-1, day-2 through day-6 and week-1 horizons. At
each point in time the model outputs a mean μ, a standard deviation σ and
a probability for each range.

MODEL
Forecasts are routed by how long the video has been observed:
- day-1: LightGBM trained on our own 2-minute series, 24h target;
- week-1 before 48h: the 24h forecast times an empirical ratio measured on
3 videos;
- from 48h: a decay LightGBM trained on public Kaggle data;
- below 2h no model is considered reliable.
LightGBM gives point estimates: σ is estimated from residuals by
observation age, with a floor that grows with the share of the horizon
still unobserved. Range probabilities are derived from μ and σ.
I am looking for experience with uncertainty and calibration on gradient
boosting models: quantile regression, conformal prediction, probability
calibration, Brier score and log-loss.

DATA AND COLLECTION
- Collection starts automatically within minutes of a new video being
published and stops exactly 168 hours later.
- View counts are recorded from the fifth minute after publication up to
the market horizon, every 40 seconds to 2 minutes.
- Market order books for every range every 2 minutes, plus executed trades
and settled outcomes.
- Range thresholds already verified against the original market questions.
- Complete videos collected with the current pipeline: 4, from 25 July to
5 September 2026. Earlier history is available but often incomplete.
The 2-minute observations are strongly autocorrelated: the unit of
analysis is the event, not the individual sample.

CODE
The pipeline is frozen at 13 July 2026 under a reference tag and has not
been modified since. Read access to code and data will be provided, under
an NDA if preferred.

WORK REQUESTED
1. Criteria fixed in writing before seeing any results: value at the
horizon, metrics and thresholds, on the frozen code version.
2. Audit: bias of μ, coverage of σ and calibration of the probabilities
using reliability, Brier and log-loss, against a naive forecaster and a
uniform distribution. Results stratified by observation time, with
attention to 0-2h and 2-6h, where errors are largest.
3. Calibration: recalibration of μ, σ and the range probabilities, with
numerical parameters that can be applied to the model, reported in
full.
4. Impact on expected value: how much the remaining probability error
translates into expected-value error, by observation-time band.
5. Final report with an outcome (pass, fail, insufficient data),
limitations, uncertainty and a procedure repeatable on future videos.

TIMELINE
There is no fixed end date. Collection continues with every new video,
roughly one every 14 days, and the project ends when collection ends. It
is up to you to decide when the data is enough.

OUT OF SCOPE
Software development, infrastructure, integration with the execution
system. Payment is tied to delivery of the work, not to betting results.

IN YOUR QUOTE, PLEASE INCLUDE
- a single fixed price, possibly with payment milestones;
- how many complete videos you consider necessary, and why;
- whether and how the price changes with the number of videos;
- what you can deliver right away on the current data;
- examples of work on probability calibration or probabilistic
forecasting.

Kind regards