Estimation & Fermi Interview Questions

Sample Answer

Roughly $10^{14}$ to $10^{15}$ total parameters across the industry. You should decompose this into three tiers: large quant firms (roughly 50 firms running thousands of models each, many with millions to billions of parameters from deep learning and NLP), mid-tier funds and banks (roughly 500 institutions with hundreds of simpler models averaging thousands to millions of parameters), and the long tail of smaller shops (5,000+ firms with dozens of lightweight models). The top tier dominates: 50 firms $\times$ 2,000 models $\times$ $10^9$ average parameters already gives $10^{14}$, and the inclusion of large language models now deployed for research and trading signals pushes the upper bound further.

Sample Answer

You could decompose top-down by starting from total US electricity consumption and estimating the data center share, or bottom-up by counting data centers and estimating per-facility costs. The top-down approach wins here because data center electricity share is a well-anchored figure (about 2 to 3% of US consumption), giving you a quick sanity check. Your tree should be: total US electricity consumption (about 4,000 TWh/year), times data center share (2.5%), times average commercial electricity price (roughly $\$0.07$/kWh), yielding $4{,}000 \times 10^9 \times 0.025 \times 0.07 \approx \$7$ billion. The key sub-components to flag are PUE (power usage effectiveness, typically 1.2 to 1.5) which inflates total facility power beyond just compute, and the price variation between regions where hyperscalers negotiate cheaper rates.

Sample Answer

You could approach this top-down using wealth distribution percentiles, or bottom-up by estimating home equity and financial assets across income brackets. The top-down approach wins here because publicly known anchors make it fast: there are about 130 million US households, and roughly the top 10 to 12 percent have a net worth above $1 million, giving you approximately $130M \times 0.11 \approx 14$ million households. This aligns well with Federal Reserve Survey of Consumer Finances data. State your assumptions on what counts as net worth (including primary residence or not) to show the interviewer you think precisely about definitions.

Sample Answer

Let me reason through this step by step. Start with the major consumer segments: investment banks, hedge funds, asset managers, corporate treasury desks, and some academic institutions. There are roughly 5,000 hedge funds globally with an average of maybe 3 to 5 terminals each, giving about 20,000. Large investment banks number around 20 to 30 globally, each with hundreds to low thousands of terminals, contributing roughly 50,000 to 80,000. Asset managers, insurance firms, pension funds, and corporates might add another 150,000 to 200,000. Summing these segments, you get roughly 300,000 to 350,000 active terminal subscriptions worldwide, which is consistent with known Bloomberg subscriber estimates in that range.

Sample Answer

This question is checking whether you can segment a nascent market and apply reasonable per-user economics rather than relying on a single top-down number. Start with roughly 330 million US residents, about 250 million adults, and surveys suggest around 15 to 20 percent have owned crypto, so roughly 40 million. Of those, maybe 25 percent are active traders in a given year, yielding about 10 million active retail traders. If the average active trader transacts $5,000 per year and pays an effective fee rate of 0.5%, that is $25 per trader in fees, giving a total addressable market of $$10{,}000{,}000 \times 25 = \$250 \text{ million per year}$$. Adjust this up or down based on your assumptions about average trade size and fee compression from competition.

Sample Answer

Start by computing total daily notional: $1{,}200 \times 4.7\text{M} = 5{,}640\text{M}$, which you can get by doing $1{,}200 \times 5 = 6{,}000\text{M}$ and subtracting $1{,}200 \times 0.3 = 360\text{M}$. Now 0.8 basis points means $0.8 \times 10^{-4} = 8 \times 10^{-5}$. Multiply: $5{,}640{,}000{,}000 \times 8 \times 10^{-5} = 5{,}640 \times 80 = 451{,}200$. So daily revenue is roughly $\$451{,}200$, and you should flag that rounding $4.7$ to $5$ early would have given $\$480{,}000$, a useful sanity check on your precision.

Sample Answer

This question is checking whether you can anchor to known perfect squares and refine with linear approximation under pressure. You know $88^2 = 7{,}744$ and $89^2 = 7{,}921$, so $\sqrt{7{,}850}$ lies between 88 and 89. The gap is $7{,}921 - 7{,}744 = 177$, and $7{,}850 - 7{,}744 = 106$, so you interpolate: $88 + \frac{106}{177} \approx 88 + 0.6 = 88.6$. The true value is approximately $88.599$, so $88.6$ nails it.

Sample Answer

Start with what the interviewer is really testing: this question is checking whether you can chain simple estimates without losing track of scale, and whether you sanity-check intermediate results. Chicago has about 2.7 million people, roughly 1 million households, and maybe 5 to 10 percent own a piano, giving around 50,000 to 100,000 pianos. Each piano is tuned once or twice a year, and a tuner can do about 4 per day across 250 working days, so roughly 1,000 tunings per tuner per year. That gives you $\frac{100{,}000}{1{,}000} = 100$ tuners, so the order of magnitude is $10^2$. If you ever got a number like 10,000 tuners, you should flag it: that would mean one tuner per 270 people, which is clearly implausible for such a niche profession.

Sample Answer

Give the rule of thumb, then the exception: the standard move is to start from annual searches and divide down. Google processes roughly 8 to 9 billion searches per day, which gives $\frac{8 \times 10^9}{86{,}400} \approx 100{,}000$ searches per second. But here, time-of-day variation matters because global usage peaks when the Americas and Europe are both awake, so peak seconds could be 2 to 3 times higher, around $2 \times 10^5$ to $3 \times 10^5$. The answer is solidly in the $10^5$ range: hundreds of thousands per second, not millions.

Sample Answer

The standard move is to bound from above using the total addressable profit pool in that instrument's market. But here, capacity constraints matter because a single instrument's daily volume and bid-ask spread cap how much alpha you can extract before you become the market. For an upper bound, take the instrument's average daily volume (say $1B notional), assume you capture at most 5 to 10 bps of that flow as profit, giving roughly $500K per day or about $125M per year, so $200M is already pushing past the ceiling. For a lower bound, consider that a decent strategy on one liquid instrument might capture $10 to $50M annually. Your $200M estimate likely overestimates either the fraction of volume captured or the per-trade edge, and you should vary those two parameters independently to see which one drives the answer past the upper bound.

Sample Answer

Get this wrong in production and you present a staffing or market-sizing recommendation that is off by an order of magnitude, destroying your credibility. You should rank your assumptions by their multiplicative impact on the final answer: the number of pianos per household, the fraction of households with pianos, tuning frequency, and tunings per tuner per day. Vary each by $\pm 50\%$ and see which one swings the output most. Typically the fraction of households owning a piano is the most uncertain (estimates range from 3% to 10%), and a 3x error there cascades directly into a 3x error in the tuner count. Halving that fraction from, say, 10% to 5% and adjusting tunings per day from 4 to 5 likely closes most of the 10x gap.

Sample Answer

Get this wrong in production and you massively undersize your quoting infrastructure or overallocate memory to dead strikes. The right call is to count expiries first: AAPL has weeklies out about 6 weeks, monthlies out roughly 6 months, and quarterlies/LEAPs adding maybe 8 more, totaling around 20 active expiry dates. For each expiry, strikes are listed in $2.50 or $5 increments across a range of roughly $\pm30\%$ around spot, giving you about 50 to 80 strikes per expiry, and you double that for puts and calls. That yields roughly $20 \times 70 \times 2 \approx 2{,}800$ actively quoted contracts, though only a fraction carry meaningful open interest.

Sample Answer

Saying billions sounds reasonable but breaks under the reality that per-share fees are fractions of a cent. Saying thousands doesn't work because even modest volume at any positive fee rate exceeds that. That leaves a millions-per-day range. US equity markets trade roughly 10 to 12 billion shares daily across all venues, and NYSE captures about 20 to 25% of that, so roughly 2.5 billion shares. At an average net capture of about $\$0.0005$ to $\$0.001$ per share (after rebates), daily transaction revenue lands around $\$1.5$ to $\$2.5$ million, which annualizes to roughly $\$400$ to $\$600$ million, consistent with ICE's reported exchange segment revenues.