Reverse stress testing the DeepSeek shock

On Monday January 27, Nvidia (NVDA) stock declined by 17%, or by more than $590 billion in capitalization – the largest drop in market cap for a single company in US market history. In turn, the S&P 500 sank 1.46%. 

Back in July 2024, as Nvidia’s share of the US indices grew, many clients asked us if our models could provide any foresight on what might happen to the market if Nvidia were to suffer a sharp decline. In response, we published ‘The impact of Nvidia on market returns’, where we shocked an index or a factor to estimate the effect it might have on an asset or a portfolio – a stress test in reverse.

In that piece, we noted that Nvidia was responsible for almost 18% of the Size factor return YTD, through June 24, 9.4% of the Momentum factor’s return and nearly 20% of the Profitability factor return.

Given the rout in NVDA on Monday, we thought our inverted stress test might be worth another look. This time we used the newest Axioma Equity Factor Risk Model (US5.1) Short Horizon and Trading Horizon models, in both their fundamental and statistical variants.

Figure 1: NVDA Return Shock: -16.97% (Jan. 27, 2025)

Source: Axioma Equity Factor Risk Model (US5.1)

Clearly, each of the models overestimates the Nvidia effect on the S&P 500 on Monday. While the broader AI trade reversed across industries such as Independent Power Producers, Electric Equipment, Tech Hardware, Communications Equipment, Electric Utilities, and even Construction, these are not large components of the S&P 500. Outside of the Semiconductors (-0.65%) and Tech Hardware industries (-0.17%), no other industry factor (with the Market Intercept added back) contributed more than 10 basis points in losses. 

When we look at what the model had predicted at the open on January 27 (using model data from January 24) and compare it to the actual factor contributions from January 27, we see the below. 

Figure 2: US5.1 Short Horizon Fundamental Model Shocks by Factor

Source: Source: Axioma Equity Factor Risk Model (US5.1)

Some of the Industry betas to NVDA appear too high, such as Software (0.023) and Tech Hardware (0.105), but the betas for Size (0.039) and for Semiconductors (0.183) appear to be good estimates, as the predicted shock returns are quite close to the actual factor contributions to the index return. With Software, the estimated beta to NVDA is modestly positive at 0.023, and the predicted shock contribution is second only to Semiconductors at -41 bps. However, the industry contribution was flat. Even with specific returns included, the total contribution from the 20 software stocks making up 10.5% of the index was just -21 bps.

Figure 3: Sector Weight Return Contributions

Source: S&P 500, Axioma Equity Factor Risk Model (US5.1)

Stocks like Oracle and Microsoft which are very closely associated with the big AI models in both infrastructure and deployment, did do poorly but other firms like Salesforce, Workday, and ServiceNow – that may actually benefit from cheaper to license AI technology in their platforms – did well.

Note that some of the factors with negative estimated betas to NVDA such as Banks and Capital Markets have negative predicted shock returns. This is because we are adding back the Market Intercept return to each of the Industry factors to make them total, as opposed to active returns over the broad market represented by the model estimation universe.

While Nvidia made back almost half its losses on January 28, we should be wary of how precarious the valuations of the stocks in the “AI Trade” might be in industries only tangentially related to the GPU manufacturers themselves, and be prepared for even larger drawdowns upon any news that disrupts the narrative.