LLMs for Business Valuation

What this hour looks like

A working session: you audit machine-read financials, design a forecasting prompt, and run a valuation simulation — then we walk the solutions.

From cash flow to firm value, in two steps

A valuation is just two ideas: figure out the cash the business throws off each year, then add up that cash in today's money.

A tiny growth error is a huge valuation error

The terminal value divides by the gap between the discount rate and the growth rate. When that gap is small, nudging the growth rate even slightly swings the whole valuation.

Make the model narrate before it numbers

A good forecasting prompt forces four stages of reasoning before any number appears — so the logic is visible and checkable.

The LLM read a 10-K — but one field is mis-extracted

You pointed an extraction model at a mid-cap industrial's FY2023 10-K. It returned this JSON (USD millions). A classic failure hides inside: it copied CapEx from the cash-flow statement without the parenthesis that marks an outflow, and flattened a footnote.

Catch the bad field, then price the damage

The extraction figures are on the previous slide. Work through four tasks:

1. Audit the extraction. Which textual features of a 10-K make CapEx, parenthesised negatives, or footnoted figures error-prone for an LLM? Name the automated check (e.g. a sign/units cross-check against the cash-flow statement) you would add to catch it.
2. Compute FCFF. Using the audited figures, compute FCFF with the effective tax rate (infer it from income tax / EBT). This \(\text{FCFF}_0\) feeds the case study.
3. Propagate the error. Recompute the DCF enterprise value (WACC = 10%, 5-yr growth = 6%, terminal growth = 2.5%) with the mis-extracted CapEx vs. the corrected value. How big a valuation error does one bad field cause?
4. State a rule. At what extraction-confidence level would you block an automated valuation from reaching a portfolio manager without human review?

Build a CoT forecasting prompt from MD&A guidance

You are building a valuation agent for the same industrial company. Its MD&A states:

Your tasks:

1. Write the four-stage CoT reasoning (macro → industry → firm → synthesis) a well-designed prompt should elicit.
2. Identify at least one potential internal inconsistency in the guidance that a consistency-checking step should flag.
3. Specify the JSON output format for Year 1–3 FCF forecasts (base, upside, downside, justification).

What the sensitivity surface looks like

Before drawing thousands of WACC and growth combinations, look at the sensitivity heatmap for a real company — it shows exactly why the distribution shape matters.

Replace point estimates with a distribution

Instead of guessing a single WACC and growth rate, you draw thousands of plausible combinations and watch the whole range of enterprise values that results. The base cash flow is your \(\text{FCFF}_0\) from Problem 1.

Draw the parameters

First, sample the uncertain inputs and clip terminal growth safely below the discount rate.

import numpy as np

def monte_carlo_dcf(fcff0, g_exp_mean, g_exp_std, wacc_mean, wacc_std,
                    g_term_mean, g_term_std, n=5, M=5000, seed=42):
    rng = np.random.default_rng(seed)
    wacc   = rng.normal(wacc_mean,   wacc_std,   M)
    g_term = rng.normal(g_term_mean, g_term_std, M)
    g_exp  = rng.normal(g_exp_mean,  g_exp_std,  (M, n))

    # Clip terminal growth below WACC -> finite Gordon value
    g_term = np.minimum(g_term, wacc - 0.005)

Each draw is one possible future. The valuation loop follows on the next slide.

Value each draw, collect the distribution

For every draw, grow the cash flows, discount them, add a terminal value, and store the enterprise value.

    ev = np.zeros(M)
    for m in range(M):
        fcfs = [fcff0]
        for t in range(n):
            fcfs.append(fcfs[-1] * (1 + g_exp[m, t]))
        pv_fcf = sum(fcfs[t]/(1+wacc[m])**t for t in range(1, n+1))
        tv = fcfs[n]*(1+g_term[m])/(wacc[m]-g_term[m])
        ev[m] = pv_fcf + tv/(1+wacc[m])**n
    return ev

# Use FCFF0 from Problem 1
ev_sim = monte_carlo_dcf(fcff0=???, ...)

Three questions to interrogate your output

1. Question A — where does the uncertainty come from? Compare the P10 and P90 enterprise values. What fraction of total EV uncertainty is terminal-value uncertainty vs. explicit-period FCF uncertainty? (Hint: re-run with \(\sigma_g = 0\) to isolate the explicit period.)
2. Question B — how sensitive is the headline number? At what terminal growth rate \(g\) does the median EV roughly double relative to the Problem 1 base case? (Hint: solve \(TV \propto 1/(\text{WACC}-g)\) for the target \(g\).)
3. Question C — does the market agree? A competitor is acquired at 12× EV/EBITDA. With EBITDA = EBIT + DA = $623M, does the comparables EV fall inside the Monte Carlo P10–P90 range? What does the gap (or its absence) say about pricing?

Fill in this table

When is the DCF unreliable, and how do you report it?

One dropped sign, a 3.2× overstatement

1. The audit. CapEx, parenthesised negatives, and footnoted figures are the classic LLM extraction traps — the model drops the sign/parenthesis or flattens a note. Fix: a verification sub-agent that cross-checks each figure's sign and units against the cash-flow statement and flags disagreements.
2. FCFF (corrected, CapEx as outflow). Effective tax rate \(\tau = 85/368 = 23.1\%\); \(\Delta\text{NWC}=65\), giving \(\text{FCFF}_0 = \mathbf{219.3}\) M.
3. Propagated error. If the LLM treats CapEx as \(+240\) (sign dropped), FCFF jumps to 699.3 M — a 3.2× overstatement that flows straight into the DCF (corrected EV ≈ 3,672 M vs. a grossly inflated figure). A single mis-extracted field, not the discounting, breaks the valuation.
4. The rule. Block any automated valuation whose extraction confidence is below threshold (or whose sign-check fails) from reaching a portfolio manager without human review.

The firm-stage reasoning — and the tension to flag

Five things to carry out of this session

1. FCFF is not a field in any filing. It is derived from EBIT, DA, CapEx, \(\Delta\)NWC, and the effective tax rate. Each input can carry LLM extraction error, and errors propagate multiplicatively.
2. CoT separates narrative from number. Force the model to commit to a macro/industry/firm story before generating numbers — inconsistencies become detectable (Wei et al., 2022).
3. Terminal growth dominates the uncertainty. A 100-bps error in \(g\) yields a 15–18% EV error. Report Monte Carlo percentile ranges, not just point estimates.
4. Comparables and DCF are complements. A large gap signals either mispricing or a flawed assumption — investigate, don't average.
5. Production pipelines need hallucination guards. Validate every ticker; trace every figure to a specific accession number. One hallucinated comparable can distort the whole multiple analysis.

Decompose the EV variance, then check the theory

The LLM forecasts the explicit growth path (Problem 2's CoT step), so its forecast error is the uncertainty the model injects. How much of the valuation spread does that error explain?