I've created a spreadsheet that appears to accurately predict deaths in the short term based on an estimated 24 day lag time between confirmed cases and deaths.
This simple mathematical function indicates that over the course of the next month we could see a peak in deaths that rivals or exceeds the carnage from last year, even in nations with good vaccination rates such as the US and the UK.
I'm attaching plots for those nations with traces for the confirmed cases and the actual versus predicted deaths. This is no magic model. It's dirt simple and robust, with no state variables or feedback, and apparently also amazingly accurate.
I fine-tuned the lag time by evaluating the match between predicted vs actual deaths by eye. I also added two 'anchors' to better align the lag time to the early pandemic before testing became widespread when people were dying in hospital waiting rooms with zero advance notice of confirmed case. The lag time ramps up linearly from zero to maximum based on those two 'anchors'.
I'm attaching two plots each for these two nations, one with the full pandemic time series and another zoomed in from last winter's peak through today. The markers are set to the cases and predicted deaths that apply to today's prediction, so that you can see where we currently are on this trajectory.
According to my spreadsheet, this coming carnage is already baked in based upon the confirmed cases to-date. It's as if these people are already dead and just don't know it yet.
The principle unknown are how much vaccination and Omicron are improving the case fatality rate in the very short term. It is possible that a dramatic shift in the case fatality rate could perturb the predictive capability of the function. Since it relies upon the lag time of 24 days, there's a 48 day window where abrupt shifts in trends could perturb the result. News reports I've seen don't indicate that either vaccinations or the penetration of Omicron have dramatically changed the outlook in the short term. I'm expecting that this coming surge of deaths is at least partially real, and that my predictive function is keeping up with the trends. Improvement in case fatality rate over the past two years as treatments and vaccines were rolled out don't appear to have dramatically perturbed the historical performance of this predictive function. Delays in testing and reporting seem to be more of a factor, as indicated by abrupt excursions in the US data during/ following the holidays and also during/following the January 6th insurrection.
Recent deaths data in the US and the UK indicate that the uptick in deaths has already begun. This increase is clearly visible in the raw deaths data in my plots, doesn't show up in my symmetrical weighted 13 day running average deaths trace because there's a 6 day processing delay inherent in that averaging function and it's nearly a week behind this uptick. I chose that averaging function because it maximally preserves signal integrity while keeping the averaged trace time-aligned to the raw trace and also maximally flattens the weekly reporting cycle without overly suppressing detail in the traces. This allows the predictive function to closely follow the actual deaths with good fidelity and keeps the plots easily discernible while preserving as much information as possible.
I would have posted this data sooner if it was ready. I've been developing this spreadsheet since the beginning of the pandemic and it wasn't until this week that I finally got it working to the level where I felt confident in the results. I haven't seen anything similar anywhere else. The complicated epidemiological models don't provide such an accurate short-term prediction and the one at IHME isn't automatically updating. With my formula, I just download the latest data and the result is instantly available. It doesn't require any tweaking.
I'm using the Johns Hopkins Covid-19 cases/deaths time series database and the Our World In Data Covid-19 testing database. There's no mathematical magic and no super genius involved in this spreadsheet. Compared to what epidemiologists routinely do with their modeling, it's trivial. I'm amazed at how well it works. I wasn't expecting it to conform so closely to reality, but apparently when the pandemic data is good, so is the predictive capability of my spreadsheet.
I'll post updates as I improve the spreadsheet with more nations. I haven't added the two 'anchors' for other nations yet and I've recently rewritten the entire spreadsheet from scratch to clarify and simplify it. I'm hoping to eventually add it to GitHub where I'm getting my data from. My first version from nearly two years ago wasn't useful and was poorly received by the experts. Maybe now it's nearly ready for prime time?
This simple mathematical function indicates that over the course of the next month we could see a peak in deaths that rivals or exceeds the carnage from last year, even in nations with good vaccination rates such as the US and the UK.
I'm attaching plots for those nations with traces for the confirmed cases and the actual versus predicted deaths. This is no magic model. It's dirt simple and robust, with no state variables or feedback, and apparently also amazingly accurate.
I fine-tuned the lag time by evaluating the match between predicted vs actual deaths by eye. I also added two 'anchors' to better align the lag time to the early pandemic before testing became widespread when people were dying in hospital waiting rooms with zero advance notice of confirmed case. The lag time ramps up linearly from zero to maximum based on those two 'anchors'.
I'm attaching two plots each for these two nations, one with the full pandemic time series and another zoomed in from last winter's peak through today. The markers are set to the cases and predicted deaths that apply to today's prediction, so that you can see where we currently are on this trajectory.
According to my spreadsheet, this coming carnage is already baked in based upon the confirmed cases to-date. It's as if these people are already dead and just don't know it yet.
The principle unknown are how much vaccination and Omicron are improving the case fatality rate in the very short term. It is possible that a dramatic shift in the case fatality rate could perturb the predictive capability of the function. Since it relies upon the lag time of 24 days, there's a 48 day window where abrupt shifts in trends could perturb the result. News reports I've seen don't indicate that either vaccinations or the penetration of Omicron have dramatically changed the outlook in the short term. I'm expecting that this coming surge of deaths is at least partially real, and that my predictive function is keeping up with the trends. Improvement in case fatality rate over the past two years as treatments and vaccines were rolled out don't appear to have dramatically perturbed the historical performance of this predictive function. Delays in testing and reporting seem to be more of a factor, as indicated by abrupt excursions in the US data during/ following the holidays and also during/following the January 6th insurrection.
Recent deaths data in the US and the UK indicate that the uptick in deaths has already begun. This increase is clearly visible in the raw deaths data in my plots, doesn't show up in my symmetrical weighted 13 day running average deaths trace because there's a 6 day processing delay inherent in that averaging function and it's nearly a week behind this uptick. I chose that averaging function because it maximally preserves signal integrity while keeping the averaged trace time-aligned to the raw trace and also maximally flattens the weekly reporting cycle without overly suppressing detail in the traces. This allows the predictive function to closely follow the actual deaths with good fidelity and keeps the plots easily discernible while preserving as much information as possible.
I would have posted this data sooner if it was ready. I've been developing this spreadsheet since the beginning of the pandemic and it wasn't until this week that I finally got it working to the level where I felt confident in the results. I haven't seen anything similar anywhere else. The complicated epidemiological models don't provide such an accurate short-term prediction and the one at IHME isn't automatically updating. With my formula, I just download the latest data and the result is instantly available. It doesn't require any tweaking.
I'm using the Johns Hopkins Covid-19 cases/deaths time series database and the Our World In Data Covid-19 testing database. There's no mathematical magic and no super genius involved in this spreadsheet. Compared to what epidemiologists routinely do with their modeling, it's trivial. I'm amazed at how well it works. I wasn't expecting it to conform so closely to reality, but apparently when the pandemic data is good, so is the predictive capability of my spreadsheet.
I'll post updates as I improve the spreadsheet with more nations. I haven't added the two 'anchors' for other nations yet and I've recently rewritten the entire spreadsheet from scratch to clarify and simplify it. I'm hoping to eventually add it to GitHub where I'm getting my data from. My first version from nearly two years ago wasn't useful and was poorly received by the experts. Maybe now it's nearly ready for prime time?