With “#stockmarketcrash” trending on Twitter / X pre-market yesterday morning, and as a counterpart to “SPY 572… 22 points (and a paradigm shift?) away“, in which we explored the SPY’s upside prospects, here we use the VecViz Dashboard platform to explore the downside prospects of the SPY. We consider the 1 month forward time horizon initially, and migrate down to the 1d time horizon1.
Before we begin, please know that this isn’t a bearish “market call”. The V-Score for SPY closed today (9/9/24) at +5 on a -12 to +12 scale (improved from +1 from our last note), and we try to limit directional commentary beyond that. See the Dashboards page for out of sample V-Score back-test performance for time horizons going out to 1 year.
Consider the distribution of Support as depicted by the Vector Strength Histogram
Let’s start with the Vector Strength Histogram. Every ticker’s is somewhat different and changes over time. The SPY’s presently displays roughly four tiers of Vector Strength to the downside: (a) the current price of 546 down to 438, (b) 438 down to 357, (c) 357 down to 226 and (d) 226 and below. The top tier,”a”, possesses at approximately 40% of the total Vector Strength to the downside, which we equate with the concept of “Support”. Thus, it is not surprising that the Vector Model’s (VM’s) EDB, the “Expected Down Body”2, represented in darkest blue, is contained entirely therein, across all time horizons. In contrast, Sigma 95D3 (the pink line) doesn’t pierces through it until near the end of the one year forward period, while the VM 99D (lightest blue) isn’t contained by it for a single day4.
The table below details the price probability percentiles displayed above, by forward horizon date.
Get context for the price probability %tiles by viewing their history.
The VM clearly sees much wider tails to the upside and downside than Sigma for the 21d time horizon5. This dynamic began in early 2024, as displayed in the chart below, and was last seen in early 2022, though not to this pronounced an extent, with even the 95d level for the VM now well below Sigma’s 99D level.
Identify the best Vector Set to provide context for the price probability percentile of interest.
Let’s see how the VecViz framework justifies these 99%tile VaR price levels. Clicking on the “VS” to the left of the table displayed above reveals the strongest Vector Set (“vs”) that envelopes each of these prices. We see that for both the VM and Sigma’s 99%tile 21d (or 1 month) VaR it is vs3. Interestingly, the VM’s Expected Down Body for the 21d forward horizon is also best supported by vs3. We display vs3 below.
Consider the “VecLevel” that the price of interest resides at within that Vector Set and what that signifies, taking into consideration VecDates and VecEvents.
With a core enveloping the trend line connecting the Apr2012 and Jul2024 extended down to the Oct2022 bottom, vs3 is quite wide. The VM’s 21d EDB resides in the upper center, at 509, near the 0.618 VecLevel. Sigma’s 99% VaR of 491 resides near the true center (the 0.5 VecLevel). Not much controversy there, with either level – just “mean reversion” within a larger trend that incorporates so many diverse VecEvent dynamics it is hard to argue against its continued relevancy, vague though it is.
In contrast, the VM’s 99D of 425 resides near the bottom of the core, (the 0.0 VecLevel). The decline to the Oct2022 bottom (to which the 0.0 level is linked) was largely inflation / interest rate concern driven. Thus, returning to the 0.0 level signifies the application of the Apr2012-Jul2024 growth trajectory with little improvement from Oct2022 with regard to inflation / interest rate concerns, or, alternatively, some improvement of those inflation / rate concerns offset by some decline relative to that longer term growth trajectory. The latter seems more correct than the former, but both seem tenable, particularly if the excitement around AI proves to be premature.
But isn’t the VM’s 99%tile 1 month VaR level pretty extreme, even for 99% VaR? A decline to 425 represents a decline of 22%. Such a decline in the SPY has been met or exceeded in only 0.25% of all rolling 21d periods since January 1997 (calculated offline, intend to add this functionality as a feature in some future release). Notably, about half those instances observations occurred when coming off the all time high of early 2020 that preceded the Covid sell off. July’ 2024’s all time high in the SPY is still fairly recent. So, perhaps 425 is a reasonable 21d 99% VaR level, but it would be helpful to get cross-validation from another Vector Set.
Consider additional strong, good fitting Vector Sets as well, to the extent available.
Let’s consider vs7, the best fit for the VM’s 10d 99% VaR level (which is just a couple points above the 21d level we have been discussing). We examine vs7 below. Note that there are >200 vector sets for a typical ticker, so vs7 isn’t necessarily materially “weaker” than vs3, and the VM’s 99d level may sit more squarely within it, making it well worth considering.
We see vs7 envelopes the trend line connecting the May2015 to Sep2018 top, extended down to the Apr2018 bottom, and the VM’s 10d 99%tile sits solidly at its center. Though vs7 is narrow, it has captured the SPY’s variation with only three brief, though large exceptions: the Covid bottom, the speculative euphoria peak of Jan2022, and the AI excitement driven rally of 2024. Notably, from mid 2022 through mid 2023, prior to ChatGPT and commitments of big tech to invest heavily in AI, the SPY was firmly in vs7.
In the context of vs7, the 424-426 area seems like a reasonable stressed downside level. However, the time horizon of 1 month, let alone 2 weeks (10d) seems a bit short, even for 99% VaR. There is substantial support displayed on the Vector Strength Histogram, displayed, near the head of this article, between the current price and 424-426 area. Can the SPY really traverse it to the downside that quickly? Some examination of the composition of this support is necessary to defend the validity of the 10d and 21d time horizons ascribed to these VM 99% VaR levels. To what extent is it associated with VecEvents or VecDates that seem vulnerable to losing relevance in a 99% VaR scenario?
Filter the Vector Strength Histogram to Reveal Important Support and Resistance Dependencies, and perhaps additional pertinent Vector Sets.
A major difference between vs7 and vs3 is that vs7’s anchoring tops and bottoms all occurred well before ChatGPT / the onset of the AI “gold rush”. Including a weakening of sentiment concerning the prospects for AI as part of a 99% VaR scenario seems reasonable in the current market. Thus, we filtered the Vector Strength Histogram to exclude all Vector Sets linked to the “AI gold rush” Vec Event. Could such a weakening in sentiment happen over 2-4 weeks? Perhaps – the excitement certainly built quickly, and we are headed into Q3 earnings season. Given the general dampening of spirits that would likely ensue were the AI thesis to take a substantial hit, we also filtered out all “leveled up” Vectors from all Vector Sets (i.e. VecLevels > 1.0). The resulting Vector Strength histogram is displayed below.
The chart above shows that the first area of any multi-vector set support is at 518, where we have confluence of vs1 (its 1.0 level, the top of its core) and the 0.382 level of vs8, with the 0.618 level of vs31 nearby at 515. In the table below the Vector Set study area you can see that 518 is roughly consistent with the EDB or “Expected Down Body” as well as Sigma’s 95% VaR for the 10d forward time horizon.
We depict vs8 in bold on the histogram6 and in the Vector Set Study area as well. You can see it reflects the Mar2020 bottom/Jan2022 top/Oct2022 bottom, a recent, though pre-ChatGPT time frame. There is limited history for this vector set, but downside moves to the 0.00 level have quickly followed when price moves below the 0.382 level. The 0.00 level for this channel is presently at 456, which corresponds to the VM’s 99%tile VaR for the 1d forward time horizon.
Consider the Vector Sets that filtering revealed, or the lack thereof.
Let’s finish with a discussion of this 1d VM VaR price level of 456. Conceptually, its position within vs8 arguably reflects policy neutral growth for the Mar2020-Oct2022 period from a “black swan” associated bottom (defining Covid as a “black swan”). Why policy neutral? The Mar20200 bottom reflected optimism about huge fiscal and monetary stimulus that would be forthcoming, whereas the Oct2022 bottom marked significant pessimism about the withdrawal of/ reduced capacity for such stimulus). But we have recovered (not without loss or scars) from the Covid Black Swan, and we are not in the midst of another “black swan”. Thus, without having consulted additional Vector Sets, a new Black Swan seems necessary to justify this price move, particularly in one day.
From a historic precedent perspective, 456 would be a decline of -16.5% from yesterday’s close, an unprecedented 1d move for the SPY since January 1997. To our knowledge the max 1d decline in the SPY during this period was -10.9%. The decline on Black Monday of 1987 was far greater at -20.5%, though that was pre “circuit breakers” on trading. Along those lines, a -16.5% move on the SPY (and hence, presumably, the S&P 500) would be more than enough to trigger a 15 minute pause in stock trading, as I understand the SEC’s guidance on such matters. Finally, not sure how applicable the comparison is to the SPY, but all this said, the Nikkei declined 12.4% just over a month ago.
In order to reach 456 on the filtered Vector Set Histogram above we would have to go through the core sections (0.0 to 1.0) and in some cases, the upside sections (1.382 to 2.236) of vs1, vs2, and vs4, down to their “leveled down sections” (-1.236 to -0.382). Not going to display them all here for sake of (blog load-ability), but these Vector Sets are anchored by tops and bottoms associated with the recovery from the oil / shale bust bottom of February 2016, the Financial Crisis bottom of March 2009 and the Eurozone Crisis bottom of October 2011, respectively. Given all the EPS growth (real and nominal) since the formation of these Vector Sets, the geopolitical significance of the events anchoring them, and the low rate environment in which they are anchored, a “black swan” (or at least a “grey swan”) event seems like it would be required for the SPY to level down in them to the 456 level so quickly. Such an outcome would also be consistent with our discussion of vs8 in the preceding paragraphs. The likelihood of a black swan occurring on any random day is of course unknown, and their frequency on average is well, well under 1% (which on a daily scale is once every ~5 months).
See below for our out of sample 1d horizon model VaR and OaR output for SPY. The VM is at a bigger disconnect to Sigma with regard to VaR than it was even in early 2022. Thus far that has proven to be an overly conservative estimate of losses, and we expect absent some great surprise, that it will likely continue to be so.
See the bottom dashboard on the Dashboards page for a filter of the 95% and 99% VaR breakage experience and associated Return on VaR Based Capital (ROVBC) for an equally spaced sampling of our out of sample test, covering both the VM and Sigma for the January 20, 2022 to present time frame. The results presently available for the 21d horizon across all ~150 tickers covered and then for just the SPY are displayed below.
These tables show that the VM delivers 99% VaR (and OaR) breakage for the 21d horizon close to the 1.00% target, and in line with Sigma, across all tickers, but for the single ticker SPY it, like Sigma, did not experience any breaks for the 19 dates considered (for which there was 21d’s of forward price history to evaluate breakage upon).
Conclusion
Hopefully this case study has demonstrated: (1) VecViz is focused on cognitively engaging risk analytics (2) with VecViz Dashboards risk analytics can be much more engaging than just VaR numbers on a screen, supported only by mathematical theory or black box AI, (3) there has to be a better way to identify support and resistance than manually drawing channels, or viewing an unranked, raw computer generated deluge of them, sans any descriptive info such as the VecDates and VecEvents VecViz provides. (4) Vector Model analytics are much more nuanced than Sigma, though they aren’t necessarily more accurate over every time horizon for every ticker (5) the output of any model can be layered upon the Vector Strength Histogram and its cognition can benefit from getting mapped to the best fitting, strongest Vector Set (6) the VecEvent functionality could be enhanced if it was populated by VecEvents provided by you or your favorite analyst team.
- A similar, though longer exercise could be done with longer time horizons, up to 252 days (1 year). ↩︎
- The Vector Model probability weighted average of all prices between the current price and the 95% VaR level (95d). Included along with its counterpart, EUB, or Expected Up Body, in the “Vector_BodyFrcst” item in the Vector Set Study area. ↩︎
- “95D” and “99D” are 95% and 99% Value at Risk or VaR levels, respectively. They are the prices for which the estimated probability of the ticker price residing below them on the horizon end date is 5% and 1%, respectively. Note that their counterparts, “95U” and “99U” are 95% and 99% “Opportunity at Risk” or OaR levels, respectively. ↩︎
- Illustrating the Vector Model’s incorporation of “jump” volatility dynamics typically available only in stochastic volatility models ↩︎
- As the first Vector Strength Histogram and the table that follows it indicate, this dynamic of the Vector Model seeing more prospective volatility than Sigma diminishes with regard to VaR over longer VaR time horizons. That said, it tends to increase with regard to OaR over longer OaR time horizons. ↩︎
- by clicking on it when in the dashboard ↩︎