@
View attachment 410901 View attachment 410902
In an effort to make a simple, but compelling forecast of plug-in EV market share, I have boiled things down to a random walk on a logit scale. May sound complex, but it actually makes very few assumptions. I examine market share from 2012 to 2018, computing the annual increase in logit market share (log(S/(1-S)). So I've got now 6 data points. I compute the mean and sample standard deviations. I make the assumption that future logit differences have the same distribution as was operative in the past. Thus, the future is modeled as a random walk with a certain mean and standard deviation. I can than predict the mean and variance (inclusive of parametric uncertainty) of the logit random walk. From this I can convert back into the market share scale, tracing out the mean path and a 90% predictive envelope.
It is the predictive envelop that I have not really explored before, but it is critical for understanding just how much uncertainty there is in a forecast that is properly conditioned on historical observation. If another modeler were to put forward a forecast that is substantially outside of the envelope, I would know that they are inserting information into there analysis that is not consistent with what has actually been observed since 2012. This does not mean that such a forecast is "wrong", but merely that it represents a substantial departure from historical trends. This departure may well originate in the modeler's own imagination about what make speed up or slow down the gain of market share in the future. For example, vehicle autonomy could alter the environment for EVs and ICE in ways that were not present in 2012 to 2018. Then again, Tesla has led with AP which may in fact give their cars a competitive advantage already witnessed in recent past. So these matters are largely a judgment call. I view my job as a modeler as to reveal what the data is telling us and to try not to assert my own opinions as an overlay to the data. This is the virtue of a truly statistical model over a judgmental model.
You can see the fruit of this simple analysis in the charts above. I have presented this with the y-axis as either on the nominal (market share) scale or the log market share scale. In the former, we see a typical logistic curve surrounded by the predictive envelope. Note that this envelope is widest around the year 2028. Yes, this is actually where we have the greatest uncertainty! Notice also that the envelope passes through the 50% mark between the years 2026 and 2030. Anything that might substantially accelerate or delay EV market dominance had better happen by 2025, otherwise it is just too late to make much of a difference. So if we imaging that autonomy will speed it up or "lack of public charging" (Come on, BNEF, you're better than that!) will slow it down, those things need to come into play within the next 5 years or it just won't impact he timing of market dominance
But some will look at that chart and think, "How can those tiny little historical observations blow up into such a big effect? I just can't believe that." This of course is the problem people have with intuiting exponential (or logistic) growth. So the whole forecast will strike them as a fanciful extrapolation. This is why I also present the exact same data with a log scale for market share. What is striking in the log scale is that one can see how the historical growth is strongly linear (or rather, log-linear). The pattern has been remarkably linear, and this is precisely why a random wake with drift is a compelling model. One is invited to question what exactly could take log market share off this strongly linear path. Indeed, one must see this clearly to understand why only very strong forces would really be able to knock it off course. For example, the oil crash of 2014-16 hardly makes a dint in the historical trend. To be sure it is there, but it is such a minor effect that one must look very closely for it. Of course, we know in the long run market share cannot exceed 100%, so ultimately the line must level out asymptotically. But notice that the bend does not really make much of difference until after EVs have dominated the market. Up through about 2025 EV growth will not be distinguishable from exponential growth.
In reality all forecasting is extrapolation. But if I must extrapolate, I prefer to extrapolate from data more so than from opinion. The trend is clear while the window to substantially alter the trajectory is narrow. When the data is bending to a different trajectory, I will gladly change my opinion. But for now the data are not showing signs of slowing, if anything the path is mildly speeding up.
Now let me make some predictions. 2019 PEV share comes in between 2.7% and 3.7%. Believable? How about 2020 between 3.8% and 6.1%, or 2022 between 7.4% and 14.8%. These may seem fairly wide, but not so wide as to be without consequence. Consider that BNEF is predicting on 10M EVs sold in 2025 or share of 10%. My model puts 2025 between 19% and 43%. So BNEF is already 2 standard deviations below my lower bound, which is about 2 standard deviations my mean of 29%. So BNEF is seriously bending the curve down in ways to which historical data does not bear witness.
Or let's back test this. In April, 2018, BNEF forecast 2018 to come in at 1.56M or about 1.67% share. The actual was 2.018M or 2.12% share. Let's what my method would have predicted using just 2012 thru 2015 data (a sample size of just 3). My 2015 forecast of 2018 would have centered on 2.2% with a 90% predictive interval from 1.3% to 3.5%. Yeah, a lot of uncertainty, but nailed it. My 2016 forecast of 2018 centered on 1.8% ranging from 1.3% to 2.5%. This was a little more pessimistic, but less uncertain. Then the 2017 forecast of 2018 centered around 2.0% ranging from 1.7% to 2.4%. So the predictive envelope nicely closed in on the actual. Meanwhile, BNEF low forecast was ruled out of the predictive interval by the time it was made using 2017 as the last historical datum. Presumably, BNEF's forecast had that benefit of highly granular data, their proprietary data, and a multi-industry team of analysts. Surely with all that going for them, they should have been able to produce a forecast with much less uncertainty, but in fact a sample size of 5 historical observations could have alerted them to the possibility that their prediction was high improbably.
So I am not at all saying that my very simple model is the best. That's not the point. The point is we need simple challenger models that don't assume much but capture the uncertainty of the historical record to tell us when our fancy forecasts are bending to improbable conclusions, overburdened with too much complexity, to much granularity, and to much opinion. If any modeling group out there wants to avoid embarrassing themselves with precious predictions, they would do well to stay with a predictive envelope such as I have constructed. This is what we call a challenger model. And to the rest of us, if we want to avoid being lured in by "credible" forecasts from well-funded organizations, we do well to have a few simple challenger models of our own.