After publishing preliminary results of my analysis of the Mediazona data, I sent an email to the Mediazona team, asking to comment on it. I received a very helpful response from Dima Treshchanin, which I summarize below.
He pointed out that during the first period of the war (2022-2023), the time between death and publication of obituary (or some other notification) was typically two weeks or a month with c.20% cases taking longer. Beginning in 2024, this time delay began increasing, first to six months and more and, subsequently, even longer - a year, a year and a half.
The fundamental reason for the lengthening delay is that the front has been moving slower and the kill zone has become wider, making evacuation of the bodies of the deceased more challenging. (PT: this also explains the disparity in the UKR/RUS body exchanges.)
Another reason for delay in reflecting this information in the database is a growing backlog in processing obituaries. It is currently 29,000 records, of which approximately one-half are unique obituaries (whereas the volunteers, who help Mediazona with this, can process only 5-6 thousand per month without sacrificing quality). Furthermore, finding the precise date of death often takes a year or longer (many obituaries don't give it and the date has to be discovered using other kinds of information).
Dima's conclusion is as follows:
Therefore, we can confidently count the number of dead in the first half 2026 at best in late 2028 or after the end of the war, whichever comes first. If the frontline finally stops moving or the kill zone expands further, even end of 2028 is an optimistic estimate.
We also don't believe in the 30,000-person death toll per month, but that doesn't mean we believe Russia's losses have somehow decreased. There is nothing to indicate this: no change in tactics, no reduction in the intensity of combat operations, no approach by commanders to somehow protect personnel.
Basically, now it's the same meat grinder as in late 2024-2025. The only difference is the seasonal factor: the Russian army does indeed suffer heavy losses in the winter months, with a "lull" in summer. We attribute this both to the fact that infiltration tactics are slightly more effective in the summer (soldiers are hidden by vegetation) and to the leadership's desire to achieve KPIs by certain dates (end of the year, certain political events etc).
I expect that 2025 might be as high as 2022-2024 combined, with no significant changes in early 2026.
What this means for the analysis is that I cannot assume stationarity (as I did in my post a week ago). My statistical model needs to reflect the fact that the time delay between death and data entering the database is lengthening. Let's do this and see how it affects the results.
First, I do a simple modification of the previous method: instead of using the data for the whole 4.5 years of war, I use only the last 2 years (from July 2024 to July 2026). Because I now have only two years of data, this calculation shows what is the predicted curve will look like two years from now (July 2028). This is OK, because the main question how relative counts are changing and this approach should give a reasonable approximation for that. Otherwise the method is the same.
Here's the result:
(As a reminder: the solid brown curve is the actual data, while the blue band shows the spread of possible values, according to the model.)
It's not very different from the previous prediction. But I don't really believe it because it assumes that time delay in reporting did not change during the past two years, while we know that it did.
If we want to incorporate this continually growing time delay into the method, I don't see any recourse but fitting a theoretical curve to data. I used the negative exponential with time-varying coefficients (for the formula, see the technical appendix below). Again, this approach predicts what the counts will be like two years in the future (in July 2028). Here is the result:
You can see now that this method predicts a less pronounced decline after the peak on Day 1008 (28-11-24) and the range of predicted values (the blue band showing the min-max spread) increases faster than in previous projections. This seems realistic to me. What's not realistic is the increasing trend after Day 1400 (January 2026). It is due to the statistical model overestimating early rates of growth. But it's not worth tweaking the model further because the 2026 data simply do not give us a solid basis for making forecasts (as was explained by Dima above).
So, perhaps my best guess about what the trajectory will look like in July 2028 is this:
Between December 2024 and December 2025 the trend is down, while after January 2026 it can be anything. We now wait until Summer 2028 to find out whether this statistical model's predictions will be similar to the data that will accumulate by that point.
So, what do I think now that I've done these analyses ? First, the Mediazona data, unfortunately, are not as informative about recent trends in casualties (that is, for 2026) as I hoped. Still, the time I spent on this analysis was not wasted. You cannot tell how informative data are until you subject them to a formal analysis. If your analysis incorporated all important sources of uncertainty, then you will get a quantitative answer in the width of the estimated uncertainty interval. This is what we saw - in all figures blue band explode as we get closer to the present.
Second, in my estimation Russian fatalities have been declining from the peak in Winter of 2024/25. For 2025 this conclusion is supported by all projections based on the Mediazona counts. For 2026, Mediazona numbers are not helpful, but I still don't expect a rising trend. This "guesstimate" is not based on the Mediazona numbers, but rather comes from the Attritional Warfare Model (AWM). If the model is correct, then we are currently in the dynamic regime when the size of the Ukrainian army is declining (see Is Ukraine-Russia War Entering the End Game?) and so is its capacity to inflict casualties. Note, this is not a prophecy, but a scientific prediction ( Scientific Prediction ≠ Prophecy). I don't know whether it's true or not; we will find out in 2028 or after the war is over. That's also when we will see how accurate the AWM predictions are, and in what ways the model will have gone wrong.
Returning to the original question, with which I started last week's post: is the current wave of reports in the mainstream media about Russia's impending defeat a psyop, or is it based on data ? I think it is clear that the exaggerated claims of huge Russian casualties have no empirical basis (see also what Dima said on this question above).
On a broader question, several people asked me why I devote so much space in this Substack to the War in Ukraine. This is a good question that requires a lengthy response, so I plan to answer it in my next post.
Technical Appendix
For statistically minded, here's the curve that I fit to the Mediazona data ("mzdat"):
fit <- nls(delta ~ a0 + a1*week^theta + b*exp(-interval/c), data=mzdat, start = list(a0=0, a1=0.001, b=20, c=30, theta=1), algorithm = "port")
where delta is the rate of increase in weekly counts, week is time in days from the beginning of war, interval is time in days between the week of death and the week of report, and a0, a1, b, c, and theta are parameters. The model fits the data quite well, with the coefficient of determination, R-sq = 0.801.