I hope this is the best place to ask a question like this and will not be to vague for this type of forum. I am hoping for some corrective guidance with an issue I am having.

I work for a non-profit that has asked me to try and update a formula used on our monthly dashboard provided to our executives. The formula is used to estimate how many donors we will have at the end of the month. A donor is defined as someone who passed away, met tissue donation requirements and family consented to donation.

Our previous formula was a simple formula which more or less was (Total Current Donors / Day in Month) * Days in Month, or average daily donors * days in month.

Where I am running into the most problems is dealing with the huge swings we can experience in death, or even deaths with potential for donation. Our goal is within the last week of the month be within 1% of our ending total, but I keep over fitting my final calculation where it doesn't translate month to month.

My issue is my formula isn't best designed to deal with huge fluctuation in deaths impacting the possible potential during that period. Our report is currently run every 6 hours, and during some periods we run into 0 potential deaths, and at other points 40 potential deaths and 15 donors. I have tried averaging many of our variables we consider impactful on referrals we receive, but I am heavily impacted at the end of the month due to volatility which usually calms down but not until about 3 days till the end of the month.

I come from limited stats work with my stats work coming from my undergrad degree years ago and I struggle to find materials on dealing with variables with unknown rates or even predictability. I was hoping for some recommended best practices for dealing with unknown, or variables themselves that deal with a lot of 'noise' making it hard to predict.

I know there is no best answer, but any push in the right direction would be extremely valuable and if this is the wrong forum for this question, I apologize.


1 Answer 1


What you have is a time series of counts, so you could look into models for that (search this site!). You also mention covariables, can you detail which ones? You should consider augmenting the forecasted value with a prediction interval, so the uncertainty (which might be large) is communicated.

The large day-to-day variability of the predicted value reminds me of how my android phone operates. The various apps are ordered in the order of "most used", but android calculates that order based only on very recent data, so the order fluctuates wildly, so the ordering is noisy and of little help. They should have used some kind of exponentially weighted average instead, so the oldest data is forgotten only gradually.

So that is where I would have started, look into some simple time series forecasting methods. If you can post a link to some historical data we could look into it, or if that is impossible at least some plots.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.