DEV Community: Jeroen Bouma

Understanding Risk Exposure Across Major Indices using the Finance Toolkit

Jeroen Bouma — Tue, 21 Jul 2026 14:22:00 +0000

In the week of March 31 to April 6, 2025, the S&P 500 ETF fell 5.85%. That alone is not unusual; weekly moves like that happen most years. What is unusual is what the rest of 2025 looked like around it: a year that otherwise behaved close to normal, with one week that did not. That single week is enough to push the S&P 500's kurtosis for the year to 26, roughly six times its typical reading in a calmer year like 2023. Volatility alone would not tell you that. Kurtosis does.

That is the case for looking past plain volatility when measuring risk. The Finance Toolkit's risk module covers Value at Risk, Conditional Value at Risk, maximum drawdown, the Ulcer Index, GARCH volatility, skewness, and kurtosis, each describing a different shape of risk that a single standard deviation number flattens into one figure. This article runs all five across six major indices: the S&P 500 (SPY), Nasdaq 100 (QQQ), Dow Jones (DIA), Russell 2000 (IWM), MSCI EAFE developed markets (EFA), and MSCI Emerging Markets (EEM). The source code for every calculation is on GitHub. The MCP server documentation lives here.

Setting Things Up

Start by installing the Finance Toolkit:

pip install financetoolkit

Then import the library and define a universe of tickers. The example below uses six major indices.

from financetoolkit import Toolkit

indices = Toolkit(
    tickers=["SPY", "QQQ", "DIA", "IWM", "EFA", "EEM"],
    api_key="YOUR_FMP_API_KEY",
    start_date="2019-01-01"
)

Get your FMP API key at jeroenbouma.com/fmp. The free plan covers five years of history, enough to span the period used here.

Value at Risk and Conditional VaR: How Bad Could a Bad Week Get?

Value at Risk (VaR) at the 95% confidence level answers one specific question: in the worst 5% of weeks, how much do you lose? Conditional Value at Risk (CVaR) answers the harder follow-up: once you are already in that worst 5%, how much do you actually lose on average? CVaR is always worse than VaR, the question is by how much.

var = indices.risk.get_value_at_risk(period="yearly")
cvar = indices.risk.get_conditional_value_at_risk(period="yearly")

Which returns (Value at Risk):

Year	SPY	QQQ	DIA	IWM	EFA	EEM
2019	-1.2%	-1.6%	-1.2%	-1.5%	-1.1%	-1.5%
2020	-3.2%	-3.5%	-3.2%	-3.5%	-2.6%	-2.6%
2021	-1.3%	-2.0%	-1.4%	-2.3%	-1.4%	-2.0%
2022	-2.6%	-3.5%	-2.0%	-2.8%	-2.3%	-2.3%
2023	-1.4%	-1.7%	-1.1%	-1.7%	-1.4%	-1.5%
2024	-1.3%	-1.9%	-1.1%	-1.9%	-1.5%	-1.7%
2025	-1.7%	-2.1%	-1.6%	-2.0%	-1.3%	-1.6%

And Conditional Value at Risk:

Year	SPY	QQQ	DIA	IWM	EFA	EEM
2019	-2.0%	-2.4%	-2.0%	-2.4%	-1.8%	-2.2%
2020	-5.5%	-5.5%	-6.0%	-6.6%	-5.3%	-5.7%
2021	-1.9%	-2.7%	-1.8%	-2.9%	-2.0%	-2.6%
2022	-3.4%	-4.3%	-2.8%	-3.6%	-3.0%	-3.0%
2023	-1.6%	-2.0%	-1.4%	-2.4%	-1.9%	-2.1%
2024	-1.9%	-2.7%	-1.5%	-2.9%	-1.9%	-2.2%
2025	-2.8%	-3.4%	-2.4%	-3.1%	-2.0%	-2.5%

IWM carries the worst or second-worst CVaR almost every year, including 2024 at -2.9% when most other indices were calm, a reminder that small-cap stress can build even when large-cap headlines stay quiet. The gap between VaR and CVaR is the more interesting read: for SPY in 2020, VaR was -3.2% but CVaR was -5.5%, meaning the average outcome inside that worst-5%-of-weeks bucket was nearly twice as bad as the threshold itself. That gap widens in every index during 2020 and 2022, the two broadest drawdown years in this dataset, and narrows again in calmer years like 2023.

Try this with the Finance Toolkit MCP: "Calculate the yearly Value at Risk and Conditional Value at Risk for SPY, QQQ, DIA, IWM, EFA, and EEM from 2019 to 2025. Which index has the worst tail risk, and in which years does the gap between VaR and CVaR widen the most?"

Maximum Drawdown: The Worst Case Each Index Actually Lived Through

VaR and CVaR are statistical estimates. Maximum drawdown is not an estimate, it is the actual peak-to-trough loss an investor in each index lived through.

max_drawdown = indices.risk.get_maximum_drawdown(period="yearly")

Which returns:

Year	SPY	QQQ	DIA	IWM	EFA	EEM
2019	-6.6%	-11.0%	-6.9%	-9.9%	-8.5%	-13.1%
2020	-34.1%	-28.6%	-37.1%	-41.1%	-34.0%	-33.9%
2021	-5.4%	-10.9%	-6.6%	-12.5%	-7.0%	-18.2%
2022	-25.4%	-35.2%	-22.0%	-27.3%	-30.3%	-33.3%
2023	-10.3%	-10.9%	-9.0%	-18.3%	-11.6%	-14.1%
2024	-8.4%	-13.5%	-6.1%	-10.1%	-11.1%	-11.7%
2025	-19.0%	-22.9%	-16.1%	-23.9%	-14.1%	-15.0%

2020's COVID crash hit IWM hardest at -41.1%, deeper than any other index in any year of this dataset, consistent with small-caps having less balance sheet cushion and less liquid trading during a panic. DIA, by contrast, never has the worst drawdown of any year, its blue-chip composition acting as the most resilient of the six across both 2020 and 2022. The 2022 bear market flips the script on tech: QQQ's -35.2% drawdown was the worst of that year, driven by the same rate-hike sensitivity that made growth stocks the epicenter of the selloff. EEM had the worst showing in 2019 and 2021, both years with no broad global crisis, a reminder that emerging markets can drawdown on their own schedule independent of US conditions.

Try this with the Finance Toolkit MCP: "Show the yearly maximum drawdown for SPY, QQQ, DIA, IWM, EFA, and EEM from 2019 to 2025. Which index had the single worst drawdown, and which index was most resilient across all years?"

Skewness: Which Years Had Fatter Downside Than Upside

Skewness measures asymmetry. A return distribution with negative skew has a longer, fatter left tail, infrequent but severe losses outweighing the frequency of gains. Positive skew is the opposite: frequent small losses, occasional large gains.

skewness = indices.risk.get_skewness(period="yearly")

Which returns:

Year	SPY	QQQ	DIA	IWM	EFA	EEM
2019	-0.59	-0.43	-0.65	-0.36	-0.56	-0.53
2020	-0.61	-0.61	-0.52	-0.98	-1.19	-1.23
2021	-0.34	-0.27	-0.37	-0.03	-0.56	-0.20
2022	0.05	0.10	-0.04	0.09	0.37	0.72
2023	-0.04	0.10	-0.03	0.33	-0.21	-0.03
2024	-0.52	-0.41	0.04	0.05	-0.43	0.06
2025	1.52	1.31	0.79	0.33	0.39	0.20

Every index ran negative in 2019 and 2020, the two years dominated by sharp, sudden selloffs (the late-2018 spillover and the COVID crash) rather than steady grinding declines. EFA and EEM show the most extreme negative skew of the whole table in 2020, at -1.19 and -1.23, consistent with international markets taking a sharper, more concentrated hit than US large caps that year. 2022 flips to mildly positive skew almost everywhere, since that bear market ground lower week after week rather than crashing in a handful of sessions, the opposite shape of risk from 2020 despite a similarly large drawdown. 2025's positive skew across the board, led by SPY at 1.52, reflects a year of mostly steady gains punctuated by the occasional sharp upside snap-back rather than a string of small losses.

Try this with the Finance Toolkit MCP: "Calculate yearly skewness for these six indices from 2019 to 2025. Which years show the most negative skew, and what does that imply about how those losses happened?"

Kurtosis: Why 2025 Doesn't Look Like a Normal Year

Kurtosis measures how fat the tails are relative to a normal distribution, regardless of which direction. High kurtosis means more extreme weeks than a bell curve would predict, in either direction.

kurtosis = indices.risk.get_kurtosis(period="yearly")

Which returns:

Year	SPY	QQQ	DIA	IWM	EFA	EEM
2019	6.05	5.35	6.26	4.49	5.59	4.48
2020	10.06	8.68	11.29	8.53	11.60	11.00
2021	3.59	3.86	3.69	2.99	3.58	3.43
2022	3.32	3.14	3.38	2.96	4.09	6.00
2023	2.81	2.86	3.19	3.80	3.55	3.37
2024	4.70	3.90	5.71	4.77	3.61	4.20
2025	26.14	20.50	17.56	8.73	19.13	11.16

This is where the April 2025 tariff-shock week shows up most clearly. SPY's kurtosis of 26.1 in 2025 is more than double the next-highest reading anywhere else in the table, and roughly nine times its calmest year (2023, at 2.81). QQQ and EFA show the same pattern, both above 19, while IWM's 2025 kurtosis of 8.73 is the lowest of the six, since small-caps had already been grinding through elevated volatility most of the year and one more sharp week barely moved the distribution's shape. 2020 is the only other year that comes close, when every index posted double-digit kurtosis during the COVID crash. The lesson sits side by side with the skewness table above: 2025 had positive skew (more upside than downside on average) and extreme kurtosis (one week dominating the whole year's tail risk) at the same time, two measurements that look contradictory until you remember they describe different things.

Try this with the Finance Toolkit MCP: "Calculate yearly kurtosis for SPY, QQQ, DIA, IWM, EFA, and EEM from 2019 to 2025. Which year and which index show the most extreme tail risk, and what does the skewness for that same year and index suggest about the direction of that risk?"

What This Means for Portfolio Construction

No single number in this article tells the full story on its own. Volatility alone would have missed the April 2025 shock entirely, since the rest of the year was unusually calm. Maximum drawdown alone would have missed that 2020 and 2022 were the same depth of loss but completely different in shape, one a crash, one a grind. CVaR alone would have missed that IWM's tail risk shows up in quiet years like 2024 as much as in crisis years.

Putting them side by side gives a more honest picture: Dow Jones (DIA) has been the most consistently resilient by drawdown across this entire period, Russell 2000 (IWM) carries the worst tail risk almost regardless of which metric you use, and emerging markets (EEM) move on their own schedule independent of what the US indices are doing. None of that changes the diversification argument, it just specifies what each piece of a portfolio is actually buying protection against.

Tracking Macroeconomic Indicators with the Finance Toolkit

Jeroen Bouma — Tue, 14 Jul 2026 14:03:00 +0000

In 2022, Brazil's central bank had its policy rate at 13.75%. The Bank of Japan's was at -0.1%. Both countries were responding to the same global shock, a wave of post-pandemic inflation that touched nearly every economy on earth, and they responded in almost opposite ways. One was years into an aggressive tightening cycle. The other had not raised rates above zero in over a decade.

That divergence is the most interesting thing macro data shows you: not that economies move together, but how differently they move through the same event. The Finance Toolkit's Economics module pulls unemployment, GDP growth, inflation, government debt, central bank rates, and bond yields for 60+ countries going back, in some series, over a century, sourced from the OECD and the Global Macro Database. Unlike the rest of the Toolkit, none of this requires an FMP API key. It is public macro data, free to query, which also makes it one of the easiest modules to point an MCP-connected assistant at without anyone needing to set anything up first.

This article tracks five economies, the United States, United Kingdom, Germany, Japan, and Brazil, through the 2021-2023 inflation shock and the years since. The source code for every calculation is on GitHub. The MCP server documentation lives here.

Setting Things Up

Start by installing the Finance Toolkit:

pip install financetoolkit

The Economics module works standalone, no ticker, no API key:

from financetoolkit import Economics

economics = Economics(start_date="2019-01-01")

Every method below takes a countries argument. The full list of supported countries and indicators is in the documentation; this article sticks to five economies chosen for contrast rather than completeness.

Inflation: One Shock, Different Timing

The global inflation spike gets talked about as a single event, but the data shows it arrived in waves, not all at once.

inflation = economics.get_inflation_rate(
    countries=["United States", "United Kingdom", "Germany", "Japan", "Brazil"]
)

Which returns:

Year	Germany	United Kingdom	Japan	Brazil	United States
2019	1.3%	1.8%	0.5%	3.7%	1.8%
2020	0.5%	1.0%	0.0%	3.2%	1.2%
2021	3.1%	2.5%	-0.2%	8.3%	4.7%
2022	6.9%	7.9%	2.5%	9.3%	8.0%
2023	5.9%	6.8%	3.3%	4.6%	4.1%
2024	2.4%	2.6%	2.2%	4.3%	3.0%
2025	2.0%	2.1%	2.0%	3.6%	1.9%

Brazil's inflation took off a full year before the others, hitting 8.3% in 2021 while the US was still at 4.7% and Japan was in mild deflation at -0.2%. By 2022 the developed economies caught up, Germany at 6.9%, the UK at 7.9%, the US at 8.0%, but Brazil had already peaked and was on its way back down to 4.6% by 2023. Japan never had a real inflation problem by international standards; its highest reading in this entire window is 3.3%, a number that would have counted as a good year almost anywhere else.

Try this with the Finance Toolkit MCP: "Compare the inflation rate for the United States, United Kingdom, Germany, Japan, and Brazil from 2019 to 2025. Which country's inflation peaked first?"

Central Banks Respond, But Not on the Same Clock

Inflation timing explains a lot of what comes next, because central banks move in response to their own country's data, not anyone else's.

policy_rate = economics.get_central_bank_policy_rate(
    countries=["United States", "United Kingdom", "Germany", "Japan", "Brazil"]
)

Which returns:

Year	Germany	United Kingdom	Japan	Brazil	United States
2019	-0.45%	0.75%	-0.10%	4.50%	1.625%
2020	-0.50%	0.10%	-0.10%	2.00%	0.125%
2021	-0.50%	0.25%	-0.10%	9.25%	0.125%
2022	0.44%	3.50%	-0.10%	13.75%	4.375%
2023	3.63%	5.25%	-0.10%	11.75%	5.375%
2024	3.81%	4.75%	0.25%	12.25%	4.375%
2025	2.88%	4.13%	0.50%	8.31%	4.255%

Brazil started hiking in 2021, a full year ahead of the Federal Reserve and the Bank of England, and took its policy rate from 2.0% to 13.75% in eighteen months, the steepest tightening cycle of the five. By the time the Fed and the Bank of England started moving in 2022, Brazil was already near its peak. The European Central Bank (the rate driving Germany's number here) lagged furthest behind, staying negative until mid-2022 and not clearing 3% until 2023.

Japan is the outlier that matters most. The Bank of Japan held its policy rate at -0.10% through the entire inflation shock, only inching to 0.25% in 2024 and 0.50% in 2025, a complete reversal of how every other central bank in this table behaved after more than a decade of fighting deflation.

Try this with the Finance Toolkit MCP: "Show the central bank policy rate for the same five countries from 2019 to 2025. Which country hiked first, which hiked the most, and which barely moved at all?"

The Labor Market Question: Did Higher Rates Cost Jobs?

The textbook expectation is that aggressive rate hikes cool the labor market. The data only partly agrees.

unemployment = economics.get_unemployment_rate(
    countries=["United States", "United Kingdom", "Germany", "Japan", "Brazil"]
)

Which returns:

Year	Germany	United Kingdom	Japan	Brazil	United States
2019	3.0%	3.9%	2.4%	12.0%	3.7%
2020	3.6%	4.7%	2.8%	13.8%	8.1%
2021	3.6%	4.6%	2.8%	13.2%	5.4%
2022	3.1%	3.9%	2.6%	9.3%	3.6%
2023	3.0%	4.0%	2.6%	8.0%	3.6%
2024	3.4%	4.3%	2.5%	7.2%	4.1%
2025	3.2%	4.1%	2.5%	7.2%	4.4%

The US is the clearest case of a pandemic-driven spike rather than a rate-driven one: unemployment jumped to 8.1% in 2020 from lockdowns, then fell back to 3.6% by 2022, the same year the Fed started its steepest hikes since the 1980s. Unemployment barely moved after that. That is the soft landing debate in one row of a table: rates rose nearly five points and the labor market shrugged.

Brazil tells the opposite story in the most striking way. Despite running the highest policy rate of any country here for three straight years, Brazilian unemployment fell every single year from 2020 onward, from 13.8% to 7.2% by 2025. High rates did not stop the labor market from healing; whatever was driving Brazilian employment had little to do with the cost of borrowing at the margin. Germany and Japan, the two economies with the smallest rate moves, also show the smallest unemployment swings, which is closer to what theory predicts.

Try this with the Finance Toolkit MCP: "Pull the unemployment rate for these five countries from 2019 to 2025. Did unemployment rise when central banks raised rates, and which country shows the clearest soft landing?"

Government Debt: Who Paid Down the COVID Bill?

Inflation and rate hikes affect more than households and businesses, they change the math on government debt, both the cost of servicing it and the rate at which it gets inflated away.

debt_to_gdp = economics.get_government_debt_to_gdp_ratio(
    countries=["United States", "United Kingdom", "Germany", "Japan", "Brazil"]
)

Which returns:

Year	Germany	United Kingdom	Japan	Brazil	United States
2019	58.6%	85.7%	236.4%	87.1%	108.0%
2020	67.9%	105.8%	258.4%	96.0%	131.8%
2021	67.9%	105.1%	253.7%	88.9%	124.5%
2022	64.8%	99.6%	256.3%	83.9%	118.6%
2023	62.7%	100.0%	249.7%	84.7%	118.7%
2024	62.7%	101.8%	251.2%	87.6%	121.0%
2025	62.1%	103.8%	248.7%	92.0%	124.1%

Japan's debt-to-GDP ratio of roughly 249% dwarfs every other country here, more than double the United States and nearly two and a half times the United Kingdom. It has stayed in a tight band for years, which is itself notable: a decade of near-zero rates means rolling over that debt costs almost nothing, a luxury the BOJ's reluctance to hike helps preserve.

The US and UK both jumped sharply in 2020 from COVID stimulus, 108% to 132% for the US and 86% to 106% for the UK, and neither has come close to working that back down; both sit higher in 2025 than they did in 2022. Germany and Brazil show the opposite pattern: both peaked in 2020-2021 and have since declined, Germany from 67.9% to 62.1%, Brazil from 96.0% to a low of 83.9% in 2022 before drifting back up to 92.0%. Brazil's case is partly mechanical. The same high inflation that pushed its central bank to 13.75% also inflated away a chunk of the real value of its debt, the kind of side effect that does not show up if you only look at the policy rate in isolation.

Try this with the Finance Toolkit MCP: "Compare government debt-to-GDP for these five countries from 2019 to 2025. Which countries reduced their debt burden after the COVID spike, and which kept climbing?"

What This Means for the Next Cycle

Pulling these four indicators side by side for the same five countries makes a point that looking at any one of them alone would miss: there is no single global business cycle, only a global shock that each economy absorbed and is unwinding on its own schedule, shaped by its starting debt level, its central bank's tolerance for inflation, and the structure of its labor market.

Brazil moved first and is furthest along in normalizing, having hiked early, peaked early, and started easing while the US and UK were still raising. Japan is the mirror image, having barely participated in the global tightening cycle and only now taking its first tentative steps away from negative rates after more than a decade. The US delivered the cleanest soft landing in the group: a sharp hiking cycle with almost no labor market damage, though its debt-to-GDP ratio shows the COVID stimulus bill has not gone away. Germany is the closest thing to textbook here, moderate inflation, moderate hikes, and the only economy in the group steadily reducing its debt burden.

What to watch next: whether Japan's exit from negative rates accelerates as inflation proves more persistent than the BOJ expects, whether US and UK debt-to-GDP keeps climbing without consequence, and whether Brazil's early-mover advantage on rate cuts gives it room to support growth while the others are still catching up.

Competitor and Sector Analysis with the Finance Toolkit

Jeroen Bouma — Tue, 07 Jul 2026 13:33:00 +0000

In 2010, Intel generated $43.6 billion in revenue and its nearest competitor in logic chips, AMD, generated $6.5 billion. NVIDIA was a $3.3 billion company still best known for gaming graphics cards. Qualcomm was growing rapidly on the smartphone wave. Broadcom and Texas Instruments were mid-sized analog and connectivity specialists.

By 2025, the same six companies tell a radically different story. NVIDIA has grown to $130.5 billion in revenue driven almost entirely by AI infrastructure demand. Intel sits at $52.9 billion, barely changed from its 2020 peak, now reporting losses. AMD has reached $34.6 billion and crossed Intel's gross margin for the first time in the company's history.

The Finance Toolkit makes it straightforward to track this transformation through financial data, both via Python code and the MCP server for those who prefer conversational analysis. The source code for every calculation is on GitHub. The MCP server documentation lives here.

Try this with the Finance Toolkit MCP: "Let's do a deep dive in the Semiconductor industry. What trends do you see in the last 10 years? And what about the fundamentals?"

Setting Things Up

Start by installing the Finance Toolkit:

pip install financetoolkit

Then import the library and define a universe of tickers. The example below uses six semiconductor companies: Intel, AMD, NVIDIA, Qualcomm, Broadcom, and Texas Instruments.

import pandas as pd
from financetoolkit import Toolkit

sector = Toolkit(
    tickers=["INTC", "AMD", "NVDA", "QCOM", "AVGO", "TXN"],
    api_key="YOUR_FMP_API_KEY",
    start_date="2010-01-01"
)

Get your FMP API key at jeroenbouma.com/fmp. A paid plan is required to access the full 15-year history used here.

Revenue: How the Landscape Shifted

The most immediate observation from pulling 15 years of income statements is the NVIDIA revenue trajectory. Everything else looks like normal cyclical variation by comparison.

income = sector.get_income_statement()

revenue = income.loc["Revenue"][
    ["2010", "2014", "2018", "2022", "2025"]
] / 1e9

Which returns:

Ticker	2010	2014	2018	2022	2025
INTC	43.6	55.9	70.8	63.1	52.9
AMD	6.5	5.5	6.5	23.6	34.6
NVDA	3.3	4.1	9.7	26.9	130.5
QCOM	11.0	26.5	22.7	44.2	44.3
AVGO	2.1	4.3	20.8	33.2	63.9
TXN	14.0	13.0	15.8	20.0	17.7

Several things stand out. Intel peaked somewhere between 2020 and 2022, then contracted, a pattern without precedent in the company's history. NVIDIA's inflection is visible between 2022 and 2025: revenue grew from $26.9 billion to $130.5 billion in three years. Broadcom's growth looks smooth in this table but masks something important: the jump from $4.3 billion in 2014 to $20.8 billion in 2018 was almost entirely acquisition-driven, as the company absorbed Avago Technologies, Brocade, and CA Technologies in rapid succession. TXN and QCOM, the two companies least exposed to the AI compute buildout, show the flattest trajectories.

Gross Margins: The Fabless Advantage

Revenue tells you who is growing. Gross margins tell you who controls their cost structure. In semiconductors, the key structural divide is between integrated device manufacturers (companies that design and fabricate their own chips) and fabless companies that outsource manufacturing to foundries like TSMC.

gross_margin = sector.ratios.get_gross_margin()

margins = gross_margin[["2010", "2014", "2018", "2022", "2025"]] * 100

Which returns:

Ticker	2010	2014	2018	2022	2025
INTC	66.1%	63.7%	61.7%	42.6%	34.8%
AMD	45.6%	33.4%	37.8%	44.9%	49.5%
NVDA	35.4%	54.9%	59.9%	64.9%	75.0%
QCOM	68.0%	59.7%	54.9%	57.8%	55.4%
AVGO	46.2%	43.9%	51.5%	66.6%	67.8%
TXN	53.6%	56.9%	65.1%	68.8%	57.0%

The Intel margin collapse is the defining story of the table. A company that earned 66% gross margins in 2010 earns 34.8% in 2025. The proximate cause is that Intel's manufacturing processes fell behind TSMC's, first by one node generation, then two, meaning Intel's chips cost more to produce while competitors using TSMC got access to better processes at lower cost. The structural consequence is that Intel's cost-per-chip disadvantage compounds every product cycle.

NVIDIA went in the opposite direction: from 35.4% in 2010 to 75.0% in 2025. A fabless company with a near-monopoly on AI training hardware can charge what the market bears, and the market has been willing to pay very high prices for H100 and Blackwell GPUs. NVIDIA's gross margin today is higher than Intel's ever was.

AMD transitioned from a company with 33% gross margins in 2014 (a period when it was fighting for survival) to 49.5% in 2025. In 2022, AMD's gross margin crossed above Intel's for the first time. That crossover is not primarily a product story, it is a manufacturing story. AMD, as a fabless company using TSMC, gained access to better nodes while Intel's own fabs struggled.

Intel vs AMD: The x86 CPU Battle

Not every company in this sector competes with every other. Texas Instruments makes analog chips for industrial and automotive customers; it does not compete with NVIDIA for AI data center revenue. Qualcomm sells ARM-based mobile processors and has never produced an x86 chip. Broadcom targets networking and storage controllers.

The genuine competitive battle in x86 CPUs is between exactly two companies: Intel and AMD. Every server, desktop, and laptop CPU socket holds a chip made by one of them. Market share gained by one is market share lost by the other.

cpu_rivals = ["INTC", "AMD"]
years = ["2010", "2014", "2018", "2020", "2022", "2024", "2025"]

cpu_revenue = income.loc["Revenue"].loc[cpu_rivals, years] / 1e9
cpu_margins = sector.ratios.get_gross_margin().loc[cpu_rivals, years] * 100
cpu_roic = sector.ratios.get_return_on_invested_capital().loc[cpu_rivals, years] * 100

Which returns:

Year	INTC Rev ($B)	AMD Rev ($B)	INTC GM	AMD GM	INTC ROIC	AMD ROIC
2010	43.6	6.5	66.1%	45.6%	n/a	n/a
2014	55.9	5.5	63.7%	33.4%	22.8%	-16.1%
2018	70.8	6.5	61.7%	37.8%	27.0%	14.9%
2020	77.9	9.8	56.0%	44.5%	23.6%	50.2%
2022	63.1	23.6	42.6%	44.9%	10.1%	4.0%
2024	53.1	25.8	32.7%	49.4%	-10.9%	2.8%
2025	52.9	34.6	34.8%	49.5%	-0.2%	6.8%

From 2010 to 2017, Intel was dominant by every metric. AMD spent this period losing money, cutting headcount, and struggling to produce a competitive architecture. Its gross margins compressed from 45.6% to below 30% in 2015 and 2016 as it competed on price with inferior products.

The inflection came in 2017 when AMD launched the Zen architecture under Lisa Su. The first Ryzen CPUs and EPYC server chips were genuinely competitive with Intel's best products. The financial data lags the product cycle: revenue did not break meaningfully higher until 2021-2022, and gross margin recovery tracked along with product mix shifting toward EPYC.

By 2022, AMD's gross margin had exceeded Intel's, a reversal that would have seemed implausible in 2015. By 2024 and 2025, Intel's ROIC had turned negative, meaning the company was destroying capital. Intel's $5 billion investment in foundry capacity between 2020 and 2024 has not yet returned the expected margins, and the company returned a net loss in 2025.

AMD's ROIC of 6.8% in 2025 is modest, reflecting the Xilinx acquisition ($35 billion, 2022) which dramatically expanded the invested capital base. The business is generating reasonable returns on operating assets; the acquisition debt is working against the ratio.

NVIDIA vs AMD: The GPU and AI Accelerator Race

AMD and NVIDIA have been GPU competitors since AMD's 2006 acquisition of ATI Technologies. Both companies design discrete graphics chips for gaming, professional workstations, and data center compute. In 2010, the two companies were comparable in size. By 2025, the gap was $130.5 billion versus $34.6 billion.

The divergence is almost entirely explained by AI infrastructure.

gpu_rivals = ["NVDA", "AMD"]

gpu_revenue = income.loc["Revenue"].loc[gpu_rivals, years] / 1e9
gpu_margins = sector.ratios.get_gross_margin().loc[gpu_rivals, years] * 100
gpu_roic = sector.ratios.get_return_on_invested_capital().loc[gpu_rivals, years] * 100

Which returns:

Year	NVDA Rev ($B)	AMD Rev ($B)	NVDA GM	AMD GM	NVDA ROIC	AMD ROIC
2010	3.3	6.5	35.4%	45.6%	n/a	n/a
2014	4.1	5.5	54.9%	33.4%	11.6%	-16.1%
2018	9.7	6.5	59.9%	37.8%	37.5%	14.9%
2020	10.9	9.8	62.0%	44.5%	24.3%	50.2%
2022	26.9	23.6	64.9%	44.9%	32.2%	4.0%
2024	60.9	25.8	72.7%	49.4%	68.4%	2.8%
2025	130.5	34.6	75.0%	49.5%	102.6%	6.8%

Through 2020, the two companies moved roughly in parallel. AMD and NVIDIA were neck-and-neck on revenue. The ROIC numbers for 2020, AMD at 50.2% and NVIDIA at 24.3%, actually show AMD leading, a reflection of AMD's lean capital base at that moment before the Xilinx acquisition.

The separation happened in 2021 and accelerated through 2022 to 2025. The driver is not gaming GPUs, where AMD's Radeon line remains a genuine competitor for NVIDIA's GeForce range. The driver is AI training infrastructure. NVIDIA's H100 and Blackwell GPUs, running on CUDA, became the default compute substrate for training large language models. The software ecosystem, built over 15 years around CUDA, proved nearly impossible for AMD to replicate quickly with its ROCm platform.

The financial consequence shows in ROIC. NVIDIA earned 102.6% return on invested capital in 2025. AMD earned 6.8%. Both companies make GPUs. One of them has a software moat that the other does not.

AMD's position in this race is structurally harder than its Intel rivalry. Against Intel, AMD competes on x86 CPU performance, a problem of engineering execution. Against NVIDIA, AMD competes on AI software ecosystem depth, a problem of developer adoption and network effects that is slower and more expensive to overcome.

What the Market Is Pricing In

The financial history explains the divergence. Current valuations reflect where the market expects each company to go.

pe = sector.ratios.get_price_to_earnings_ratio()
ev_ebitda = sector.ratios.get_ev_to_ebitda_ratio()

current_valuation = pd.DataFrame({
    "P/E": pe["2025"],
    "EV/EBITDA": ev_ebitda["2025"]
}).sort_values("EV/EBITDA")

Which returns:

Ticker	P/E	EV/EBITDA
QCOM	34.1	14.2
INTC	neg	19.1
TXN	31.9	21.3
NVDA	63.5	55.5
AVGO	72.6	50.5
AMD	80.8	52.1

Qualcomm is the cheapest name on EV/EBITDA at 14.2x, pricing in limited AI exposure and the ARM licensing exposure risk. Texas Instruments at 21.3x is being valued as a stable analog business with cyclical revenue sensitivity.

Intel is uninvestable on P/E (the company is reporting losses) and its EV/EBITDA of 19.1x is elevated relative to the earnings power the business is currently generating. The market is either pricing in a recovery or pricing in acquisition optionality. The gross margin and ROIC data suggest recovery will take years.

NVIDIA at 63.5x P/E and 55.5x EV/EBITDA is expensive in absolute terms. Whether those multiples are justified depends on whether AI infrastructure CapEx remains elevated and whether NVIDIA maintains its software ecosystem lead. At 102.6% ROIC in 2025, the business is extraordinary; the question is the duration of that advantage.

AMD at 80.8x P/E trades at a premium to NVIDIA despite lower ROIC and a two-front competitive battle. The premium reflects expectations for ROIC recovery as the Xilinx acquisition amortizes and as AI GPU shipments scale. The execution risk is real.

Screening for Undervalued Stocks with the Finance Toolkit

Jeroen Bouma — Tue, 30 Jun 2026 13:43:00 +0000

With thousands of publicly listed companies, finding undervalued stocks by hand is impractical. A systematic screen changes that: pull valuation multiples across an entire universe, filter by multiple criteria simultaneously, and overlay profitability metrics to separate genuine value from value traps. The Finance Toolkit makes each of those steps a few lines of Python.

This article walks through a complete screening process from universe definition to a final shortlist. It uses 15 stocks across five sectors to illustrate the logic. A real screen would cast a wider net, and the Discovery module supports that too.

The source code for every calculation is on GitHub. The MCP server documentation lives here.

Setting Things Up

Start by installing the Finance Toolkit:

pip install financetoolkit

Then import the library and define a universe of tickers. The example below uses 15 stocks across technology, healthcare, consumer staples, financials, and energy.

import pandas as pd
from financetoolkit import Toolkit

screener = Toolkit(
    tickers=[
        "AAPL", "MSFT", "INTC", "IBM",
        "JNJ", "PFE", "MRK",
        "KO", "WMT",
        "JPM", "BAC",
        "XOM", "CVX",
        "META", "GOOGL"
    ],
    api_key="YOUR_FMP_API_KEY",
    start_date="2022-01-01"
)

Get your FMP API key at jeroenbouma.com/fmp. The free plan covers five years of history; a paid plan unlocks deeper history and a larger universe.

Step 1: Pulling Valuation Multiples

The first pass uses four ratios to get a broad picture of relative cheapness. P/E and EV/EBITDA are the most common entry points; P/FCF and P/B provide cross-checks that are harder to manipulate.

pe = screener.ratios.get_price_to_earnings_ratio()["2025"]
pfcf = screener.ratios.get_price_to_free_cash_flow_ratio()["2025"]
pb = screener.ratios.get_price_to_book_ratio()["2025"]
ev_ebitda = screener.ratios.get_ev_to_ebitda_ratio()["2025"]

valuation = pd.DataFrame({
    "P/E": pe,
    "P/FCF": pfcf,
    "P/B": pb,
    "EV/EBITDA": ev_ebitda
}).sort_values("P/E")

Which returns:

Ticker	P/E	P/FCF	P/B	EV/EBITDA
INTC	neg	neg	1.6	19.1
BAC	14.3	32.9	1.5	14.4
MRK	14.5	21.4	5.0	10.2
JPM	16.2	8.9	2.6	18.7
XOM	18.0	21.9	2.0	9.3
PFE	18.3	15.6	1.6	9.4
JNJ	18.8	25.5	6.2	16.0
CVX	23.0	17.0	1.5	8.9
KO	23.0	56.9	9.4	22.9
IBM	26.5	24.3	8.6	21.8
META	28.1	36.8	7.8	17.1
GOOGL	29.0	52.2	9.2	25.7
MSFT	35.5	50.4	10.5	22.7
AAPL	36.4	41.3	55.3	28.7
WMT	46.3	71.1	9.9	22.6

The spread is immediately visible. BAC and MRK trade below 15x earnings. WMT trades at 46x. INTC appears at the top of the sort only because its P/E is negative, since the company reported a net loss, which is why sorting by a single metric is dangerous without additional filters.

A few other readings worth noting. KO trades at 57x free cash flow despite a relatively modest P/E of 23x, suggesting either elevated CapEx or capital structure effects. P/B of 55.3x for AAPL reflects the reality that Apple has bought back so much equity it has almost none left, a reminder that P/B is most useful for asset-heavy sectors like financials and energy.

Step 2: Applying the Filter

A single-metric screen is easy to game by accounting choices, one-time charges, or unusual capital structures. The approach here requires two metrics to pass simultaneously: a P/E below 20x and an EV/EBITDA below 18x, with negative earners excluded.

value_candidates = valuation[
    (valuation["P/E"] > 0) &
    (valuation["P/E"] < 20) &
    (valuation["EV/EBITDA"] < 18)
].index

print(list(value_candidates))

Which returns:

['BAC', 'JNJ', 'MRK', 'PFE', 'XOM']

JPM passes the P/E filter but misses on EV/EBITDA at 18.7x, marginally above the threshold. CVX passes on EV/EBITDA (8.9x) but its P/E of 23x falls just outside the cut. Five stocks pass both filters: BAC, JNJ, MRK, PFE, and XOM.

Note that EV/EBITDA is less meaningful for banks, where the concept of enterprise value and operating earnings work differently from industrial companies. JPM at 16x P/E with a P/FCF of 8.9x deserves separate consideration using bank-specific metrics like price-to-tangible-book and return on equity.

Step 3: The Quality Overlay

Cheap on valuation is not the same as undervalued. The value trap problem is real: stocks trade at low multiples because investors expect deteriorating earnings, balance sheet stress, or structural decline. Before acting on any of the five candidates, the profitability picture needs to hold up.

roe = screener.ratios.get_return_on_equity()["2025"]
roic = screener.ratios.get_return_on_invested_capital()["2025"]
gross_margin = screener.ratios.get_gross_margin()["2025"]

quality = pd.DataFrame({
    "ROE": roe,
    "ROIC": roic,
    "Gross Margin": gross_margin
}).loc[value_candidates]

Which returns:

Ticker	ROE	ROIC	Gross Margin
BAC	9.7%	4.8%	56.1%
JNJ	35.0%	33.0%	72.8%
MRK	36.9%	28.1%	71.8%
PFE	8.8%	11.3%	70.3%
XOM	10.7%	14.8%	21.7%

The table separates the field. JNJ and MRK both earn above 28% on invested capital with gross margins above 71%. These are not cheap because the business is deteriorating. They are cheap because pharmaceutical stocks have faced broader sector pressure, patent cliff concerns (MRK), and litigation overhangs (JNJ). The underlying profitability is intact.

XOM earns 14.8% ROIC, which is solid for an energy company operating with significant fixed asset bases. The thin gross margin of 21.7% reflects the commodity economics of oil refining and distribution rather than a structural weakness. The risk here is cyclical: energy earnings compress when oil prices fall.

PFE at 11.3% ROIC is borderline. Its gross margin of 70.3% confirms the underlying pharmaceutical business generates strong economics, but the headline return metrics are depressed by the revenue reset after COVID vaccine revenues ran off. Whether this is a genuine recovery opportunity or a prolonged restructuring depends on the pipeline.

BAC at 4.8% ROIC would look like a disqualifier in an industrial context. For a bank, where assets are funded by deposits rather than equity, ROE of 9.7% is the more relevant metric. That said, it suggests the market's discount is modest rather than an obvious bargain.

Step 4: The Final Screen

Five stocks passed the dual valuation filter. The last step applies a ROIC threshold to separate the genuine opportunities from those where cheap pricing reflects structurally weak returns on capital.

final_screen = quality[quality["ROIC"] > 0.10].sort_values("ROIC", ascending=False)

Which returns:

Ticker	ROE	ROIC	Gross Margin
JNJ	35.0%	33.0%	72.8%
MRK	36.9%	28.1%	71.8%
XOM	10.7%	14.8%	21.7%
PFE	8.8%	11.3%	70.3%

BAC falls out at 4.8% ROIC under the industrial threshold, though the banking context warrants separate analysis. The remaining four represent meaningfully different risk profiles.

JNJ and MRK are the strongest combination of value and quality in this screen. Both earn well above their cost of capital, carry strong franchise positions, and trade at multiples that imply no earnings growth, which is conservative given both have significant product pipelines. XOM offers commodity exposure with decent capital efficiency and a low EV/EBITDA of 9.3x, appropriate for investors comfortable with oil cycle risk. PFE is the speculative recovery candidate: the business model is intact, the discount is real, but the recovery timeline is uncertain.

What the Screen Does Not Tell You

A quantitative screen narrows the field. It does not do the work.

Each of these names requires analysis the Toolkit can support but cannot automate: debt maturity profiles, near-term earnings catalysts, management capital allocation track records, and sector-specific risk factors. MRK's patent cliff on Keytruda is a known risk not captured in trailing ROIC. XOM's capital expenditure plans depend heavily on a commodity price path no screen can forecast.

The output of this screen is a watchlist, not a buy list. What it does efficiently is eliminate the 11 stocks where the combination of price and quality does not justify closer attention.

Try this with the Finance Toolkit MCP: "Screen 15 stocks across technology, healthcare, consumer staples, financials, and energy for P/E below 20, EV/EBITDA below 18, and ROIC above 10%. Show the valuation and quality metrics side by side."

The Finance Toolkit is open-source and available on GitHub. The Finance Toolkit MCP server is accessible directly from Claude, Copilot, Cursor, Windsurf, and Gemini for analysis without writing any code.

How Microsoft Has Changed Since Its IPO: 40 Years of Financial Data with the Finance Toolkit

Jeroen Bouma — Tue, 23 Jun 2026 13:53:00 +0000

March 13, 1986. Microsoft went public at $21 per share. Split-adjusted, that works out to roughly $0.08. The stock closed 2025 at $484, making it one of the most remarkable compounding records in market history: a return of more than 5,000x over four decades.

But the stock price alone obscures something more interesting. The Microsoft that went public in 1986 and the Microsoft that generates $282 billion in annual revenue today are not the same company. The financials show three distinct businesses operating under the same name, each shaped by a different CEO, a different product strategy, and a different relationship with capital. Pulling 40 years of data through the Finance Toolkit makes that transformation visible in a way that no headline number can.

This article demonstrates using the Finance Toolkit to analyze Microsoft's financial history. It also includes MCP prompts at the end of each section so you can run the same analysis yourself with an AI assistant instead of Python, if you'd rather skip straight to the conversation. The source code for every calculation is on GitHub. The MCP server documentation lives here.

Setting Things Up

Start by installing the Finance Toolkit:

pip install financetoolkit

Then import the library and create a Toolkit instance for Microsoft, specifying the ticker, your FMP API key, and the start date for historical data:

from financetoolkit import Toolkit

company = Toolkit(
    tickers=["MSFT"],
    api_key="YOUR_FMP_API_KEY",
    start_date="1986-01-01"
)

Get your FMP API key at jeroenbouma.com/fmp. The free plan covers five years of history; a paid plan unlocks the full 40-year dataset used here.

Revenue: Three Very Different Eras

Microsoft's revenue history divides cleanly into three phases, each corresponding to a CEO tenure: Bill Gates (1986-2000), Steve Ballmer (2000-2014), and Satya Nadella (2014-present).

income_statement = company.get_income_statement()

revenue = income_statement.loc[
    ["Revenue", "Net Income"],
    ["1986", "1990", "1995", "2000", "2005", "2010", "2015", "2020", "2025"]
]

Which returns:

Year	Revenue ($B)	Net Income ($B)
1986	0.2	0.04
1990	1.2	0.3
1995	5.9	1.5
2000	23.0	9.4
2005	39.8	12.3
2010	62.5	18.8
2015	93.6	12.2
2020	143.0	44.3
2025	281.7	101.8

The Gates era was extraordinary by any standard. Revenue grew from $197 million in 1986 to $23 billion in 2000, a 115-fold increase in 14 years driven almost entirely by Windows and Office on the back of the PC revolution. Microsoft was not really a technology company in the modern sense; it was a licensing machine with near-zero cost of reproduction and near-monopoly pricing power.

The Ballmer era is often described as a lost decade, but that framing is partly unfair. Revenue grew from $23 billion to $87 billion between 2000 and 2014. That is real growth by any measure. What failed was not the business but the valuation. Ballmer inherited Microsoft at the height of the dot-com bubble, when the market was pricing in decades of compounding at rates no company could sustain. The stock went roughly nowhere for 14 years not because the business stagnated, but because the starting price was simply too high.

The 2015 net income of $12.2 billion against $93.6 billion in revenue reflects Nadella's first full year, when Microsoft took a $7.5 billion write-down on the Nokia acquisition (Wikipedia). Strip that out and the underlying business was already accelerating. By 2020, net income had grown to $44 billion. By 2025 it reached $102 billion, a number that would have seemed absurd when Nadella took over.

Try this in the Finance Toolkit MCP: "Show me Microsoft's annual revenue from 1986 to 2025, and identify the three distinct growth phases under Gates, Ballmer, and Nadella."

Profitability: Margins That Compress, Then Recover

The revenue story is straightforward. The margin story is more nuanced.

ratios = company.ratios.collect_profitability_ratios()

margins = ratios.loc[
    ["Gross Margin", "Net Profit Margin"],
    ["1990", "1995", "2000", "2005", "2010", "2015", "2020", "2025"]
]

Which returns:

Year	Gross Margin	Net Margin
1990	82.6%	23.6%
1995	89.8%	24.5%
2000	86.9%	41.0%
2005	84.8%	30.8%
2010	80.2%	30.0%
2015	64.7%	13.0%
2020	67.8%	31.0%
2025	68.8%	36.2%

Gross margins peaked near 90% in the mid-1990s, reflecting what pure software economics look like. Distributing a copy of Windows cost almost nothing once the disc was pressed. As Microsoft expanded into hardware (Xbox, Surface), services, and cloud infrastructure, the cost structure changed. Azure compute costs real money to run. So does gaming content, data center operations, and technical support at enterprise scale. By 2015, gross margins had compressed to 64.7%.

That compression has largely stabilized. Azure's gross margin has improved as the infrastructure matures and higher-margin software services (Azure AI, Microsoft 365, Copilot) grow as a share of the mix. The more important metric is net margin, which tells a different story: it has recovered to 36.2% in 2025, approaching the levels Microsoft achieved at the peak of its software monopoly in 2000. The 2015 net margin of 13% is an outlier caused by the Nokia write-down, not a structural deterioration.

The reason for the recovery is operating leverage. Revenue nearly tripled from 2015 to 2025, but operating expenses did not. Each additional dollar of cloud or subscription revenue drops to the bottom line at an increasingly favorable rate.

Try this in the Finance Toolkit MCP: "What were Microsoft's gross margin and net profit margin at five-year intervals from 1990 to 2025? Highlight where margins compressed and where they recovered."

The Balance Sheet Tells a Story by Decade

Early Microsoft had no debt. None. The business generated more cash than it could spend, and management had no desire to lever up. For most of the Gates era and the first half of the Ballmer era, the balance sheet was a cash accumulation machine.

balance_sheet = company.get_balance_sheet_statement()

snapshot = balance_sheet.loc[
    ["Cash and Short Term Investments", "Long Term Debt", "Goodwill"],
    ["1995", "2000", "2005", "2010", "2015", "2020", "2025"]
]

Which returns:

Year	Cash + Investments ($B)	Long-Term Debt ($B)	Goodwill ($B)
1995	2.0	0.0	0.0
2000	23.8	0.0	0.0
2005	37.8	0.0	0.0
2010	36.8	4.9	12.4
2015	96.5	27.8	16.9
2020	136.5	59.6	43.4
2025	94.6	40.2	119.5

The first debt appeared in 2009, not because Microsoft needed the money, but because borrowing at low rates to fund buybacks was more tax-efficient than repatriating overseas cash. By 2020, cash and investments had reached $136 billion while long-term debt stood at $60 billion. The net position was still comfortably positive, but the capital structure had become more aggressive.

The most striking shift is goodwill. It was zero for most of Microsoft's history because the company built rather than bought. That changed under Ballmer with acquisitions like aQuantive and Skype, and accelerated dramatically under Nadella. LinkedIn ($26 billion, 2016), GitHub ($7.5 billion, 2018), Nuance ($19 billion, 2022), and Activision Blizzard ($69 billion, 2023) turned Microsoft into one of the most acquisitive companies in tech (Wikipedia). Goodwill reached $119.5 billion in 2025.

The cash balance fell from $136 billion in 2020 to $94.6 billion in 2025, largely because of Activision. Long-term debt has declined from its peak of $66.7 billion in 2019 to $40.2 billion in 2025 as Microsoft pays down the acquisition-related borrowing, a sign that the balance sheet is normalising after the deal.

Free Cash Flow: The Number That Matters Most

Revenue and earnings can be managed. Free cash flow is harder to fake. It is also the number that ultimately determines what a company is worth. Not the earnings per share, not the operating income, but the actual cash that flows out of the business after maintaining and expanding the asset base.

cash_flow = company.get_cash_flow_statement()

fcf_table = cash_flow.loc[
    ["Operating Cash Flow", "Capital Expenditure", "Free Cash Flow"],
    ["1995", "2000", "2005", "2010", "2015", "2020", "2025"]
]

Which returns:

Year	Operating Cash Flow ($B)	CapEx ($B)	Free Cash Flow ($B)
1995	2.0	-0.5	1.5
2000	11.4	-0.9	10.5
2005	16.6	-0.8	15.8
2010	24.1	-2.0	22.1
2015	29.7	-5.9	23.7
2020	60.7	-15.4	45.2
2025	136.2	-64.6	71.6

Microsoft generated $71.6 billion in free cash flow in fiscal 2025, more than the entire annual revenue of companies like Netflix or Airbnb. It is the product of two converging forces: subscription and consumption revenue that scales without proportional cost increases, and a customer base that has locked itself into the Microsoft ecosystem so deeply that churn is structurally low.

There is a caveat embedded in those numbers. Capital expenditure has increased from $5.9 billion in 2015 to $64.6 billion in 2025, the vast majority of it AI infrastructure. Microsoft is making an enormous bet on data center capacity to support Azure AI and Copilot services. Whether that $65 billion in annual CapEx generates returns commensurate with its cost will be the defining capital allocation question of the next five years. For now, operating cash flow has grown fast enough to absorb it, but the ratio of FCF to operating cash flow has compressed.

Try this in the Finance Toolkit MCP: "What was Microsoft's free cash flow in 2015 and 2025? Show the operating cash flow and capital expenditure over time."

Return on Equity: What Buybacks and Leverage Changed

Return on equity sounds like a clean measure of profitability, but it is not always clean. ROE depends on the size of the equity base, which Microsoft has spent the last 15 years shrinking through aggressive buybacks and debt-funded capital return. A company can raise its ROE by buying back enough shares to reduce equity even if underlying profitability does not improve.

profitability_ratios = company.ratios.collect_profitability_ratios()

profitability_ratios.loc[
    ["Return on Equity", "Return on Assets", "Return on Invested Capital"],
    ["1990", "1995", "2000", "2010", "2015", "2020", "2025"]
]

Which returns:

Year	ROE	ROA	ROIC
1990	37.7%	30.6%	36.9%
1995	29.3%	23.1%	29.3%
2000	27.0%	20.7%	27.0%
2010	43.8%	22.9%	47.9%
2015	14.4%	7.0%	19.4%
2020	40.1%	15.1%	30.5%
2025	33.3%	18.0%	30.6%

The early Microsoft numbers are a reminder of what a capital-light monopoly looks like. ROE of 38% in 1990 with zero debt and no buybacks means the underlying business was genuinely that profitable. Those returns compressed through the 2000s as the business mix changed and the equity base remained large.

The 2015 figures are distorted by the Nokia write-down, which depressed net income and inflated the apparent collapse in ROA and ROE. By 2020 ROE had rebounded to 40%, though this partly reflects the equity base shrinking from buybacks rather than purely improved profitability. ROIC strips out the effects of capital structure and tells a cleaner story: Microsoft earns roughly 30-31% on invested capital today, which is very good but not dramatically different from what it earned in the 1990s. The business quality has been maintained; what changed is how the returns are distributed to shareholders.

Return on assets paints a similar picture. ROA of 18% in 2025 reflects that Microsoft now carries a much heavier asset base of cloud infrastructure, acquisitions, and data centers than the software company of the 1990s. The high-30s ROA of the Gates era is gone, but 18% is a strong number for a company of this size and scope.

Looking Ahead

Forty years of operating history tell you what Microsoft has been. Valuation multiples tell you what the market thinks it will be. The tables below show where Microsoft's key ratios and risk-adjusted performance metrics have stood over the past six years, alongside the Alpha and Beta figures that put the stock's behaviour in market context.

valuation_ratios = company.ratios.collect_valuation_ratios()

current_valuation = valuation_ratios.loc[
    ["Price-to-Earnings", "Price-to-Book", "Price-to-Free-Cash-Flow", "EV-to-EBITDA"],
    ["2020", "2021", "2022", "2023", "2024", "2025"]
]

performance = company.performance.collect_all_metrics()

alpha_beta_performance = performance.loc[
    ["2020", "2021", "2022", "2023", "2024", "2025"],
    ["Beta", "Alpha"]
]

Which returns:

Metric	2020	2021	2022	2023	2024	2025
P/E	38.6	41.8	24.9	38.8	35.7	35.5
P/B	14.4	18.0	10.9	13.6	11.7	10.5
P/FCF	37.8	45.6	27.8	47.2	42.5	50.4
EV/EBITDA	27.0	32.2	19.1	27.9	24.5	22.7

Metric	2020	2021	2022	2023	2024	2025
Beta	1.14	1.15	1.28	1.16	1.19	0.88
Alpha	+0.25	+0.24	-0.09	+0.33	-0.11	-0.02

At 35x earnings and 50x free cash flow, Microsoft is priced for sustained growth. The P/FCF ratio in particular looks stretched on the surface, but it needs to be read alongside the elevated CapEx. Microsoft spent $64.6 billion on capital expenditure in fiscal 2025, the vast majority of it on AI infrastructure. If that investment generates meaningful Azure revenue growth over the next three to five years, the normalized FCF picture is more attractive than the headline ratio suggests. If it does not, the multiple will prove optimistic.

Beta declining to 0.88 in 2025 is notable. For most of the Nadella era, Microsoft moved in line with or slightly above the broader market. A beta below 1 suggests the market increasingly treats it as a defensive-quality holding: a large, profitable, cash-generative business that belongs in portfolios regardless of the macro environment.

Alpha has been positive in most years, notably +0.33 in 2023 when Microsoft's AI positioning drove a 57% stock return against the S&P 500's 26%. The negative alpha in 2022 and 2024 reflects years when the market rotated toward other sectors, not fundamental deterioration. The business compounded steadily throughout.

What to watch over the next few years: the pace of Azure growth (which funds everything else), margin expansion as the AI CapEx cycle matures, and whether Activision contributes enough to earnings to justify the $69 billion price tag. The financial track record over 40 years gives Microsoft the benefit of the doubt. But at these valuations, the margin for error is narrower than it has been at any point in the company's history outside the dot-com peak.

Try this in the Finance Toolkit MCP: "Pull Microsoft's current valuation ratios and its yearly Beta and Alpha since 2020. Is the market pricing it as a growth stock or a defensive one right now?"

Access 200+ Financial Metrics with the Finance Toolkit (Python Library + MCP Server)

Jeroen Bouma — Tue, 16 Jun 2026 13:28:00 +0000

On the 6th of May, 2023, Microsoft's Price-to-Earnings ratio was reported as 28.93 by Stockopedia, 32.05 by Morningstar, 32.66 by Macrotrends, 33.67 by the Wall Street Journal, and 34.4 by Companies Market Cap. Every one of those numbers is "correct." They just use different definitions of earnings, different share counts, and different rounding. None of the providers publish the formula, so there is no way to know which one matches the calculation you actually want.

That inconsistency is why I built the Finance Toolkit: an open-source Python library where every ratio, indicator, and model is implemented in plain, readable code you can audit yourself. It covers 200+ metrics across equities, options, currencies, crypto, ETFs, indices, and macroeconomic data going back over a century, all sourced from 30+ years of financial statements. On top of the library sits an MCP server that exposes the same 200+ metrics to any AI assistant that supports the Model Context Protocol, so you (or your assistant) never have to choose between writing Python and asking a question in plain English.

The source code for every calculation is on GitHub. The MCP server documentation lives here.

Setting Things Up

The Python library and the MCP server are the same engine with two different front doors. Start with the Python side, since understanding it makes the MCP tool calls easier to reason about later.

Start by installing the Finance Toolkit:

pip install financetoolkit -U

Then import the library and create a Toolkit instance. Each section below swaps in a different ticker universe, since the point is showing the breadth of what one consistent API can pull, but the syntax never changes.

from financetoolkit import Toolkit

companies = Toolkit(
    tickers=["MSFT", "AAPL"],
    api_key="YOUR_FMP_API_KEY",
    start_date="2020-01-01"
)

Get your FMP API key at jeroenbouma.com/fmp. The free plan covers five years of history and 250 requests a day; a paid plan unlocks the full 30+ years and quarterly data, at a 15% discount through that (affiliate) link. I do provide means to provide your own data as well, see here.

From Code to Conversation: the MCP Server

Writing Python is not always the fastest way to get an answer, and getting an AI assistant to reason accurately over financial data by itself usually falls short: web search is unreliable, scraped data is inconsistent across sources (see the PE example above), and models hallucinate numbers when they cannot find them. The Finance Toolkit MCP server fixes this by giving any MCP-compatible assistant direct, structured access to the same 200+ metrics, backed by the transparent calculation methods in the library.

You do not need a local Python environment to use it. A single command configures your AI client automatically:

uvx --from "financetoolkit[mcp]" financetoolkit-mcp-setup

This supports Claude Desktop, Claude Code, GitHub Copilot in VS Code, Cursor, Windsurf, and Gemini. If you use Claude Desktop specifically, there is also a one-click MCPB bundle on the latest GitHub release that skips the terminal entirely. Either way, you will be asked for the same FMP API key used above, and the free plan is enough to get started.

Once it is running, the server groups the 200+ Finance Toolkit methods into about 21 categorical tools. You never name a function or set a parameter yourself; the assistant picks the right tool from your question and returns structured output. The depth of interpretation scales with the model: Claude Sonnet layers in qualitative reasoning on top of the numbers, while smaller models like GPT-5 mini return clean structured data without the narrative. Both work, since the server is built to support either.

Within Claude Desktop, it will look like below once setup.

Equity Analysis: Margins, Returns, and Multiples Side by Side

The Finance Toolkit was built with equity research first, which is why this is the deepest category: profitability, valuation, efficiency, liquidity, and solvency ratios, plus models like WACC, DuPont decomposition, the Altman Z-Score, and the Piotroski F-Score. A good stress test for any of these is an industry where growth rates differ wildly between names that are nominally in the same business: semiconductors.

from financetoolkit import Toolkit

chips = Toolkit(
    tickers=["NVDA", "AMD", "ASML", "TSM", "AVGO", "INTC", "QCOM", "TXN", "AMAT"],
    api_key="YOUR_FMP_API_KEY",
    start_date="2015-01-01"
)

cumulative_return = chips.get_historical_data()
pe = chips.ratios.get_price_to_earnings_ratio()["2025"]
ev_ebitda = chips.ratios.get_ev_to_ebitda_ratio()["2025"]
eps_growth = chips.ratios.get_earnings_per_share(growth=True)["2025"]
revenue_per_share = chips.ratios.get_revenue_per_share()["2025"]

Which returns:

Ticker	Cumulative Return (2015-2025)	P/E (2025)	EV/EBITDA (2025)	EPS Growth (2025)	Revenue/Share (2025)
NVDA	+42,215%	63.5x	55.5x	+146.2%	$5.26
AMD	+20,344%	80.8x	52.2x	+164.3%	$21.17
AVGO	+3,935%	72.6x	50.5x	+286.2%	$13.16
AMAT	+2,347%	29.7x	23.8x	+0.6%	$35.11
TSM	+1,981%	28.4x	18.0x	+57.2%	$23.75
ASML	+1,762%	36.9x	28.0x	+45.0%	$98.67
QCOM	+297%	34.1x	14.2x	-44.1%	$40.08
TXN	+585%	31.9x	21.3x	+4.8%	$19.37
INTC	+352%	-	19.1x	-98.75%	$10.88

The dispersion inside one industry is the whole story here: NVDA turned a 2015 position into roughly 421 times its starting value, while INTC is (compared to NVDA) essentially flat a decade later, its P/E reported as negative because it posted a net loss and its EPS collapsing 98.75% in 2025 alone. AVGO's EPS growth of 286.2% outpaces even NVDA, reflecting the VMware acquisition layered on top of its AI networking business rather than organic chip sales growth, worth separating out before reading too much into the headline number. ASML trades at the richest EV/EBITDA multiple in the group (28.0x) despite the slowest revenue-per-share base of the bunch in absolute growth terms, a premium the market assigns to its effective monopoly on the EUV lithography machines every leading-edge fab depends on.

A call with the Finance Toolkit MCP would return an answer such as below.

Technical Analysis: Is the Market Overbought?

Technical indicators work off OHLC price data, so they apply to any asset class the Toolkit supports, not just equities. RSI, MACD, the Stochastic Oscillator, and Bollinger Bands are all available, grouped under categories like momentum, volatility, and overlap. The clearest stress test for a momentum indicator is a market shock, so this one looks at European sector ETFs through the 2020 Corona crash and the year that followed.

from financetoolkit import Toolkit

sectors = Toolkit(
    tickers=["EXV9.DE", "EXV1.DE", "EXV4.DE", "EXI5.DE", "EXV3.DE", "EXH9.DE", "EXH1.DE"],
    api_key="YOUR_FMP_API_KEY",
    start_date="2019-11-01"
)
rsi = sectors.technicals.get_relative_strength_index(period="monthly")

These are iShares STOXX Europe 600 sector UCITS ETFs: EXV9.DE (Travel & Leisure), EXV1.DE (Banks), EXV4.DE (Health Care), EXI5.DE (Real Estate), EXV3.DE (Technology), EXH9.DE (Utilities), and EXH1.DE (Oil & Gas).

Which returns:

Date	Travel & Leisure	Banks	Health Care	Real Estate	Technology	Utilities	Oil & Gas
2019-12	63.3	51.4	77.0	70.4	72.7	89.7	51.1
2020-03	26.9	31.9	65.9	36.2	53.2	51.9	26.8
2020-06	30.2	26.2	70.2	39.5	59.1	59.9	21.4
2020-09	34.9	28.9	61.5	44.0	61.6	57.6	19.3
2021-01	40.7	35.4	48.4	40.7	61.0	57.9	34.5
2021-05	77.0	79.0	63.3	75.3	80.9	71.2	61.1
2021-12	61.7	79.5	79.1	74.5	83.3	65.9	76.3

Oil & Gas was the most oversold sector of the whole crisis, bottoming at an RSI of 19.3 in September 2020, well past the typical oversold line of 30, and it stayed depressed there for most of the year as travel demand collapse fed straight through to fuel demand. Banks were close behind, troughing at 26.2 in June 2020 and not climbing back above the 40 mark until November, longer than any other sector here took to recover. Health Care and Technology never came close to oversold even at the worst of the March 2020 crash, with RSI bottoming at 65.9 and 53.2 respectively, the closest thing to a flight-to-quality signal in this dataset.

The more striking part is the reversal. The same two sectors that bottomed hardest, Oil & Gas and Banks, both closed 2021 deep in overbought territory at 76.3 and 79.5, a swing of roughly 50 to 60 RSI points in eighteen months. That is the kind of regime change a single point-in-time RSI reading would never catch; it only shows up when you pull the full series.

A call with the Finance Toolkit MCP is showing what I mean:

Risk and Performance: What the Fama-French Factors Reveal About South American Stocks

Beta and alpha alone treat "the market" as the only systematic risk that matters. The Fama-French five-factor model goes further, decomposing returns into exposure to the market (Mkt-RF), company size (SMB), value versus growth (HML), profitability (RMW), and investment intensity (CMA), plus an R-squared that tells you how much of the return that combination actually explains. South American equities are a good test case, since they mix commodity exporters, banks, and a tech name that behaves nothing like the rest of the region.

from financetoolkit import Toolkit

south_america = Toolkit(
    tickers=["PBR", "YPF", "VALE", "SCCO", "ITUB", "CIB", "MELI", "GGB"],
    api_key="YOUR_FMP_API_KEY",
    start_date="2021-01-01"
)
fama_french = south_america.performance.get_fama_and_french_model(period="yearly")

Which returns, for 2025:

Ticker	Mkt-RF	SMB	HML	RMW	CMA	R²
ITUB (Itau, bank)	0.0167	-0.0023	-0.0071	-0.0116	-0.0051	0.683
SCCO (Southern Copper, mining)	0.0146	0.0000	-0.0047	-0.0117	-0.0033	0.529
VALE (Vale, mining)	0.0118	0.0042	-0.0086	0.0032	-0.0103	0.511
CIB (Bancolombia, bank)	0.0109	-0.0029	0.0024	-0.0107	-0.0101	0.340
MELI (MercadoLibre, e-commerce)	0.0087	-0.0001	-0.0005	-0.0023	-0.0049	0.122
YPF (YPF, energy)	0.0067	0.0023	-0.0027	-0.0041	-0.0008	0.192
GGB (Gerdau, steel)	0.0063	-0.0029	-0.0014	-0.0110	-0.0009	0.276
PBR (Petrobras, energy)	0.0057	-0.0017	0.0002	-0.0037	0.0009	0.169

The R-squared column is the most useful read here, since it tells you how reliable the rest of the row is. ITUB's 2025 regression explains 68.3% of its return variance through these five factors, the highest in the group and typical of a large, liquid bank whose returns track broad risk factors closely. MELI sits at the other extreme: only 12.2% of its return is explained by the model, which fits a company whose price action is driven far more by idiosyncratic growth narrative and earnings surprises than by size, value, or investment-style exposures.

That gap widens when you look at the trend across years rather than a single snapshot:

Ticker	2021	2022	2023	2024	2025
MELI R²	0.619	0.347	0.213	0.334	0.122
VALE R²	0.131	0.196	0.408	0.489	0.511
SCCO R²	0.149	0.288	0.299	0.471	0.529

MELI's R-squared has fallen almost every year since 2021, from 0.619 to 0.122, meaning the five-factor model explains less and less of its return over time as the stock has decoupled further from traditional risk factors. VALE and SCCO show the opposite pattern, with explanatory power roughly tripling over the same five years as both mining names became more tightly linked to broad market and value-style risk during the commodity cycle. None of that shows up in a beta or alpha calculation alone, which is the case for running the full factor model instead of stopping at single-factor CAPM.

A call with the Finance Toolkit MCP would return an answer such as below. Note that the LLM picked a different timeframe which is why there is a mismatch between these results and the table above.

Macroeconomic Analysis: Tracking Asia's Growth Engines

Macroeconomic data goes back over a century and covers 60+ countries: unemployment, GDP growth, inflation, trade balances, government debt, central bank rates, and government bond yields, sourced from the OECD and the Global Macro Database. It is available through the economics module on a Toolkit instance, or as a fully standalone module if you only need macro data.

from financetoolkit import Economics

economics = Economics(start_date="2020-01-01")

gdp_growth = economics.get_gross_domestic_product(
    countries=["India", "Vietnam", "Indonesia", "China", "South Korea"],
    growth=True
)

Which returns:

Country	2020	2021	2022	2023	2024	2025
India	-1.2%	18.9%	14.2%	9.6%	10.1%	10.4%
Vietnam	4.4%	5.5%	12.5%	7.1%	10.4%	9.6%
Indonesia	-2.5%	9.9%	15.4%	6.7%	7.6%	7.7%
China	3.5%	11.7%	5.0%	4.6%	4.5%	6.4%
South Korea	0.9%	7.9%	4.6%	3.3%	5.5%	4.0%

India has held growth above 9.5% every year since the 2020 contraction, while China has settled into a slower mid-single-digit pace after its 2021 reopening spike. South Korea, the most mature economy in this group, tracks closest to developed-market growth rates throughout. None of this required scraping a single government website or reconciling conflicting definitions across sources, which is the same problem this whole project started from.

A call with the Finance Toolkit MCP would return an answer such as below.

What's Next

This only covers four of the roughly 21 categories the MCP server groups its tools into. The same approach extends to options Greeks, fixed income yield curves, ESG scores, revenue segmentation by product and geography and a Discovery module for screening across 80,000+ tickers. A Portfolio module that loads your own positions from a spreadsheet is also available albeit only through the Python library due to the added complexity. All of it is documented here, and all of it is queryable the moment the MCP server is connected.