In many ways, 2023 is shaping up to be a year of AI hype. Generative AI and large language models have become hot buzzwords, and have many wondering just which jobs aren’t in any danger.
I’m willing to put a stake in the ground here and admit that I don’t expect AI to cause significant employment disruptions in the asset management world. I’ll happily explain my thinking—but first, let’s define the debate.
A lot of quantitative funds and traders already do something similar to what AI might be able to do—they use algorithms and machine learning techniques to either execute trades under optimal market conditions, or generate actual buy/sell decisions and portfolio allocations. For this universe of investors, AI shouldn’t have a major impact because the algorithms are already in place and playing significant roles in investment decisions. Ultimately, the algorithms are calibrated and controlled by people, who make the value judgements around different parameters that constrain model outputs. In terms of how effective these techniques are, I’m not an expert—I know smart people who place a lot of weight on quant algorithms, but I’ve never understood just how this relates to AI, nor felt compelled to invest based on these models.
What I want to talk more about is the fundamental investment world—investing in stocks based on the relationships between their prices and their intrinsic values. AI doesn’t provide much help here, because valuing a company requires nuance and interpreting conflicting trends and facts. Only a truly general AI would be able to compete with humans in a bottom-up, company-specific approach to valuation—and this general AI doesn’t yet exist.
That said, I was curious how major AI tools would perform. So I ran a few tickers through Bing AI, Microsoft’s version of the product, to see what approaches the AI would take to valuing companies. For each company listed, I used the following prompt structure:
Write a 350-word report on the intrinsic value of Microsoft stock. The ticker is NasdaqGS:MSFT.
Let’s take a look. Note: the Bing AI output is italicized.
Also note: I (the human) do not recommend any investment action (buy, sell, or hold) on any of the securities discussed.
Microsoft: Turning on the Parent
Microsoft Corporation (NasdaqGS:MSFT) is one of the world’s leading technology companies, offering software, services, devices, and solutions. The company has been growing its revenue, operating income, net income, free cash flow, and return on equity consistently over the past decade. To estimate the intrinsic value of Microsoft stock, we can use the discounted cash flow (DCF) method, which projects the company’s future cash flows and discounts them to their present value.
The DCF method requires two inputs: the expected growth rate of free cash flow and the discount rate. The growth rate of free cash flow can be estimated based on the company’s historical performance, analyst forecasts, and industry trends. The discount rate can be calculated using the capital asset pricing model (CAPM), which reflects the risk and opportunity cost of investing in Microsoft.
Based on the data from Nasdaq.com, Simply Wall St, GuruFocus, and Wealthy Education, we can use the following assumptions for our DCF analysis:
– Free cash flow in 2021: $55.6 billion
– Growth rate of free cash flow for the next 10 years: 11.84% in 2022, 8.88% in 2027, 4.35% in 2030, and 2% thereafter
– Discount rate: 6.5%, based on a risk-free rate of 2%, a beta of 0.8, and a market risk premium of 5.63%
– Terminal value: calculated using the perpetual growth method
Using these assumptions, we can project the free cash flow for Microsoft for the next 10 years and beyond, and discount them to their present value using the discount rate. The sum of the present value of the free cash flow and the terminal value is the enterprise value of Microsoft. To get the equity value, we need to subtract the net debt (total debt minus cash and equivalents) from the enterprise value. To get the intrinsic value per share, we need to divide the equity value by the number of shares outstanding.
The DCF analysis gives us an intrinsic value per share of $245.51 for Microsoft stock. Compared to the current market price of $322.46 (as of August 24, 2023), Microsoft stock is overvalued by 24%. Therefore, based on this valuation method, Microsoft stock is not a good investment opportunity at this price level.
I figured this would be a warm-up. There’s plenty of internet commentary and coverage on Microsoft, which the model appears to draw on. And the structure of the valuation seems to make sense, although the model doesn’t give a ton of detail into their assumptions. Plus, I think it’s funny that we managed to convince Bing to label their parent company’s stock a “sell”.
That said, there are still some holes. The fact that the base year for FCF is 2021 instead of 2022 is odd, perhaps because the model chose an older source. It’s also unclear how the model is choosing among competing sources for the same information. And of course, we have to talk about the discount rate. The risk premium of the market is a bit low for my taste, given that 7% is often used as a long-term average. But what really throws me off is the risk-free rate: 2%? In this economy? I don’t understand the logic of a 2% risk-free rate when the Fed’s inflation target is 2%, and the 10-year Treasury is currently yielding 4.8%. Again, it’s unclear whether the model is properly accounting for more up-to-date information. But on the whole, Bing AI seemed to be able to pull together the pieces to create a valuation. Let’s try another one—here’s another large-cap with not quite as much popular following.
Deere & Co.: Fawning Over Recent Performance
Deere & Co. is a leading manufacturer of agricultural and construction equipment, with a global presence and a strong brand recognition. The company has been benefiting from the robust demand for its products, driven by the recovery of the global economy, the favorable commodity prices, and the increased adoption of precision agriculture technologies. Deere & Co. has also been investing in innovation, digitalization, and sustainability, which are expected to enhance its competitive advantage and long-term growth prospects.
The company reported impressive results for its third quarter of 2023, beating analysts’ estimates on both revenue and earnings. The company also raised its guidance for the full year, reflecting its confidence in the outlook. Deere & Co. generated $13.8 billion in revenue, up 12% year-over-year, and $10.2 in earnings per share, up 24% year-over-year. The company’s operating margin improved to 18.6%, compared to 16.2% in the same period last year.
Based on the company’s strong performance and growth potential, we estimate the intrinsic value of Deere & Co. stock using a discounted cash flow (DCF) model. We assume a weighted average cost of capital (WACC) of 8%, a terminal growth rate of 3%, and a free cash flow (FCF) projection based on the company’s guidance and historical trends. We obtain an intrinsic value of $462 per share, which implies an upside of 17.5% from the current market price of $386.85 per share (as of August 24, 2023).
Therefore, we recommend buying Deere & Co. stock at the current price level, as we believe it is undervalued by the market and offers an attractive opportunity for long-term investors. The company has a solid track record of delivering consistent growth, profitability, and shareholder returns, and is well-positioned to capitalize on the favorable industry trends and customer demand.
I picked Deere simply because we covered that in last week’s post, and I wanted to get a comparison. On the whole, it seems pretty similar to the Microsoft valuation—taking a base year’s profitability and extrapolating it forward with a terminal DCF model. I was impressed by the noting of factors contributing to the firm’s success—agricultural sector strength, commodity prices, etc.—some of which I also highlighted. It’s easy, but usually too easy, to just take a base year at face value, and it’s usually risky to assume the recent past continues forever. The positive language around the recent performance is reminiscent of a lot of sell-side research I’ve seen, and it makes me wonder whether that is part of the input set the model is using—which I find odd, given I’d expect a lot of that content to be paywalled; this raises even more questions around how much paywalled content the model is able to access.
My biggest concern, however, is that the model doesn’t simply stop at giving us a value. It gives a recommendation—saying that because the value it calculates represents a premium vs. the market price, we mere mortals should buy the stock. I’m not sure whether this is the machine thinking for itself or simply parroting a source it drew from—but either way, I’m curious to see how such algorithmic behavior unfolds. Financial service is highly regulated, and while I’m sure Bing AI could pass the Series 7 pretty quickly, it isn’t a licensed broker.
One possible outcome of all this is that specific AI algorithms that are tailored to these sorts of approaches eventually do get some blessing from the SEC to act as brokers. But can you trust a machine to act in your best interests if there’s no one to punish for malfeasance? And is Microsoft liable if I put all my money in Deere stock & lose it all?
Let’s move on. I want to crank up the difficulty a bit, so here’s an obscure European small-cap.
Bystronic AG: Forgetting Formulas
Bystronic AG is a Swiss company that provides solutions for sheet metal processing, laser cutting, bending and automation. The company has a market capitalization of CHF 2.3 billion as of August 24, 2023, and a trailing twelve-month revenue of CHF 1.1 billion.
To estimate the intrinsic value of Bystronic AG, we can use the discounted cash flow (DCF) method, which projects the future free cash flows of the company and discounts them to the present value using an appropriate discount rate. The discount rate reflects the risk and opportunity cost of investing in Bystronic AG, and can be estimated using the weighted average cost of capital (WACC) formula.
The WACC formula is:
WACC = E / (E + D) * Re + D / (E + D) * Rd * (1 – T)
where E is the market value of equity, D is the market value of debt, Re is the cost of equity, Rd is the cost of debt, and T is the effective tax rate.
Using the data from Bystronic AG’s financial statements and market data, we can calculate the WACC as follows:
E = CHF 2.3 billion
D = CHF 0.2 billion
Re = 6.5% (estimated using the capital asset pricing model)
Rd = 2.5% (estimated using the yield to maturity of Bystronic AG’s bonds)
T = 18% (average effective tax rate for Bystronic AG in the past five years)
WACC = 2.3 / (2.3 + 0.2) * 0.065 + 0.2 / (2.3 + 0.2) * 0.025 * (1 – 0.18)
WACC = 0.058 or 5.8%
Using the WACC as the discount rate, we can project the free cash flows of Bystronic AG for the next five years using the following assumptions:
– Revenue growth rate: 5% per year (based on the average growth rate for Bystronic AG in the past five years)
– Operating margin: 12% (based on the average operating margin for Bystronic AG in the past five years)
– Capital expenditure: 5% of revenue (based on the average capital expenditure for Bystronic AG in the past five years)
– Working capital: 10% of revenue (based on the average working capital for Bystronic AG in the past five years)
The projected free cash flows are:
The terminal value of Bystronic AG can be estimated using the perpetual growth method, which assumes that the free cash flow will grow at a constant rate forever after the last projected year. The perpetual growth rate should be lower than the long-term growth rate of the economy, and we can use a conservative estimate of 2%. The terminal value is:
Terminal Value = Free Cash Flow in Year N+1 / (WACC – Perpetual Growth Rate)
Terminal Value = -42 * (1 + 0.02) / (0.058 – 0.02)
Terminal Value = CHF -1 billion
The present value of the free cash flows and the terminal value can be calculated by discounting them using the WACC:
The sum of the present values of the free cash flows and the terminal value is the enterprise value of Bystronic AG, which is:
Enterprise Value = CHF -869.6 million
To get the equity value of Bystronic AG, we need to subtract the net debt (debt minus cash) from the enterprise value. The net debt of Bystronic AG as of December 31, 2022, is:
Net Debt = Debt – Cash
Net Debt = CHF 0.2 billion – CHF 0.1 billion
Net Debt = CHF 0.1 billion
The equity value of Bystronic AG is:
Equity Value = Enterprise Value – Net Debt
Equity Value = CHF -869.6 million – CHF 0.1 billion
Equity Value = CHF -969.6 million
The intrinsic value per share of Bystronic AG is:
Intrinsic Value Per Share = Equity Value / Number of Shares Outstanding
Intrinsic Value Per Share = CHF -969.6 million / 20 million
Intrinsic Value Per Share = CHF -48.48
Based on this analysis, the intrinsic value per share of Bystronic AG is negative, which means that the company is overvalued by the market and not a good investment.
Sheesh: even when I converted the data tables into Excel form, BingAI did not show much respect for word count. The methodology appears thorough. The reason I chose Bystronic is because the company has recently gone through a series of divestitures, meaning that past history is not necessarily relevant to the new company, and I wondered if BingAI would be able to recognize that despite a lack of attention for the stock. It appears that it can’t, since it’s just taking the last five years of data and extrapolating again. This shouldn’t be much of a surprise given the names we’ve tested already—but at least a five-year average is somewhat more nuanced than a single base year.
But the elephant in the room: a negative intrinsic value? Clearly the model missed a formula step here. A look through the tables will reveal it. Instead of deducting the incremental working capital to arrive at annual free cash flow, the model deducted the entire new balance of working capital, creating distorted negative cash flows that drove a negative stock price.
I’ve run Bystronic through BingAI a handful of times, and it never fails to disappoint. Sometimes the model claims overvalued, sometimes undervalued. One time, it got the current price of the shares completely wrong—I ran it again with the same prompt, and it got the right price. It raises questions about quality/availability of input data, and suggests a pretty inconsistent methodology. Let’s try again with an even more obscure name.
Thungela Resources, Ltd.: “In terms of data, we have no data.”
Thungela Resources Ltd (JSE:TGA) is a South African company that produces and exports thermal coal. According to a fundamental analysis by Alpha Spread, the intrinsic value of one TNGRF stock (the US listing of Thungela) is 31.737 USD, which is 76% higher than the current market price of 7.515 USD. This indicates that Thungela is undervalued by the market and has a potential for growth. However, the company also faces challenges such as environmental pressure, mine-rehabilitation costs, and volatility in coal prices.
This may have been unfair. Thungela Resources has only reported a couple years of results, and the model appears to have a pretty strong U.S. bias. It appears like a cop-out, but that may be the best the model can do. It provides a source of where they’re getting an intrinsic value estimate, and lists a few legitimate caveats to interpreting that estimate. To me, this suggests that there’s a certain level of obscurity at which the model struggles to find data—although this doesn’t make a ton of sense given the thorough incorrectness of the AI’s approach on Bystronic.
RXO, Inc.: “What’s my name again?”
At risk of this post getting a big long, I want to briefly touch on RXO, Inc. I ran this through the BingAI three times, each with the same prompt as I outlined at the start. RXO is a truck brokerage firm that spun out of XPO last year, so there’s very little clean history from a reporting standpoint.
The first try, the model’s output seemed to make sense. It made reasonable growth assumptions (5% for 5 years, 3% thereafter) along the lines of Bystronic. It did make the same working capital mistake as with Bystronic, but reversed the sign on their terminal value number to ultimately get a positive value for the stock. The model again picked a risk-free rate of 2%, which irritated me given current rate environments, but added a 1% premium for the company’s BBB- rating, which is pretty reasonable and appears based on the current yield curve. On the whole, it seemed like a good showing, just a bit wordy (over 1100 words when I’d asked for 350). So I left the prompt unchanged, and tried again.
When I clicked that button, RXO, Inc. was suddenly transformed into a cloud-based provider of healthcare solutions. It was an identity crisis that a Google search had no answers for. The AI had simply decided that a new, fictitious company was the real RXO, and that it was deeply undervalued and I should buy it. I rebooted again, hoping the model would come to its senses—but alas! It doubled down, insisting that RXO, Inc. was in fact a healthcare-tech firm, and I should buy the stock.
To me, this was the strangest case. Is it possible that RXO was just so new that there wasn’t enough Internet traffic on it to drive a unified definition of the company? I had provided the ticker and the name, so I don’t know what would have caused the model to veer off course. When I search for “RXO, Inc. cloud-based healthcare” I get hits directing me to RXO, Inc.’s website—RXO, the logistics and truck brokerage firm.
So What Does It Mean?
This is obviously a pretty motley collection of test cases—and that was my intention. On the whole, I come away from these examples with a fair bit of comfort, that equity research jobs are likely safe from AI for a while. The model delivers what I would consider to be particularly bad sell-side research: extrapolating recent trends without justification, recommending stocks as fundamentally undervalued without a compelling base to the argument, and dismissing any necessity for nuance and judgement in evaluating and pricing company-specific risks. It doesn’t provide any of the meaningful nuggets of information that good sell-side research is full of. Essentially, it can’t evaluate context and apply judgement the way a human can.
That said, a properly-calibrated tool could make use of AI techniques to greatly improve the efficiency of analysts to do research more quickly—getting a summary of long company reports, scanning an entire universe of earnings call recordings to search for mentions of a particular product of factor, etc. This harps on the theme that AI will be more of a productivity tool than a replacement for human labor altogether. In the medium-term, this strikes me as feasible, although such models would likely have to be built specifically for such purposes, as general-use chatbots and large language models—clearly—don’t seem to be cutting it.
I suppose the best way to get an AI to take over the entire security analysis process would have to start with careful limits to the input data. For instance, a value investor might calibrate a model using exclusively the writings of Benjamin Graham & Warren Buffett, and add as inputs the relevant company reports, earnings presentations, transcripts, sell-side research, and industry data that is relevant to each company being evaluated. An approach like this would likely have a firm grasp in the mechanics of the valuation process, and possibly a more consistent and thoughtful strategy to deal with judgements on a company-by-company basis. But could such a model truly replicate human discretion?
And even if it could: how would we hold it accountable? Trusting any decision to an algorithm creates risk. Of course, trusting human decisions involves risk too, but there’s a clear culprit there if something goes wrong. With AI, who gets the blame? The model developers? The firm that created the model? How can we punish nonhuman actors, or prevent rogue software engineers from creating meme stock phenomena on steroids?
That’s part of what makes AI so controversial. We have to answer whether we trust our fellow humans more than the creations of our fellow humans. We’ve never placed more trust in machines and algorithms than we currently do—but that trend has come at a cost. How we handle AI may be a step function on the order of the Internet or running water. But are we ready for it? What will it cost?
I won’t pretend the know the answers to these questions, but it certainly makes you think—certainly a lot harder than BingAI did trying to value Thungela Resources, Ltd.