10.5 · Build-along: the Kaveri model in Python

Why this matters: you have read about every piece — historicals, drivers, schedules, statements, DCF, scenarios, checks. This lesson puts them together in ~60 lines of Python that reproduce the reference valuation (₹320), extend it with a stress case, and then point the same code at a real Indian company. Type it, run it, break it, fix it. The model you understand is the one you built.

Learning objectives — after this lesson you can:

• Build a three-statement projection and DCF for Kaveri with tools/fi and reproduce ₹320/share.
• Extend it with a receivables stress case and a margin scenario, and read the balance-sheet consequences.
• Adapt the pipeline to a real NSE-listed company using fi.data.fetch_statements, and know what to check.
• Explain the simplifications the model makes and when they matter.

Prerequisites: 10.1–10.4; Python with pandas; the course tools installed (appendix) · Time: ~120 min hands-on

---

0. Setup

pip install -r tools/requirements.txt
python -m pytest tools/tests -q          # 46 passed
python tools/examples/build_kaveri_model.py

The example script is the finished build-along; the sections below reconstruct it step by step. Work in a script or notebook with sys.path including tools/ (the examples do this for you).

1. Load the historicals and prove they tie

import io, sys
sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf-8")     # ₹ on Windows consoles
sys.path.insert(0, "tools")
from fi.data import load_kaveri
hist = load_kaveri()                     # line items × FY21…FY26, ₹ Cr
print(hist.loc[["rev", "ebitda", "pat", "cfo", "cash", "total_assets", "total_le"]].round(1))
assert (abs(hist.loc["total_assets"] - hist.loc["total_le"]) < 0.05).all()

Six balanced years. For a real company this step is 10.2 and takes a day.

2. State the drivers

from fi.model import Assumptions
base = Assumptions.kaveri_base()
print(base)

Assumptions.kaveri_base() returns the reference drivers: growth 10%, 14%, 14%, 13%, 12%, 11%, 10%, 9%, 8%, 7%; EBITDA margin 13.3% → 15.5%; D&A 3.4% and capex 3.5% of revenue; NWC 25%, 23%, then 22% of revenue; tax 25.17%; dividend payout 25% of prior-year PAT; interest 8.9% on opening borrowings; 8.5% on leases; 6.8% treasury yield; term-loan repayments of ₹20 Cr a year for four years. Override anything by keyword: Assumptions.kaveri_base(nwc_pct=0.277).

3. Project the statements

from fi.model import project
proj = project(hist, base)
inc, bs, cf, checks = proj["income"], proj["balance"], proj["cash"], proj["checks"]
print(inc.round(1)); print(bs.round(1)); print(cf.round(1))
print(checks.loc[["balance_diff", "cash_recon_diff", "cash_negative"]].round(6))

What to look at, in order:

1. Checks first. balance_diff and cash_recon_diff are zero in every year; cash_negative is zero. If any is not, stop.
2. The P&L path. Revenue 1,450 → 3,667; EBITDA 193 → 568; PAT 96 → ~370 by FY36. Finance costs fall as term debt is repaid; other income rises as cash builds.
3. The balance sheet. Cash climbs from ₹32 Cr to ₹87 Cr (FY27), ₹150 Cr (FY28), ₹220 Cr (FY29) and keeps rising — the model is telling you that if the drivers hold, Kaveri becomes cash-rich, and the capital-allocation question of 05.5 becomes live.
4. The cash-flow statement. CFO jumps from ₹65 Cr (FY26) to ₹163 Cr (FY27) — the working-capital normalisation. That line is the thesis; everything else follows from it.

4. From statements to value

from fi.model import fcff_from_projection
from fi.valuation import capm, wacc, terminal_value, dcf_fcff, equity_bridge
w = wacc(capm(rf=0.065, beta=1.05, erp=0.06), kd_pre_tax=0.089, tax_rate=0.2517, debt_weight=0.10)   # 0.1219
fcff = fcff_from_projection(proj)                     # FY27…FY36: 108.5, 114.2, 124.5, … 275.6
nopat_next = inc.loc["nopat"].iloc[-1] * 1.055
tv = terminal_value(nopat_next, w, 0.055, ronic=0.18)          # 3,638.7
d = dcf_fcff(fcff.tolist(), w, tv, mid_year=True)               # EV 2,100.0
equity = equity_bridge(d["ev"], net_debt=140.5, leases=18.2)    # 1,941.3
print(round(equity / 6.07))                                      # 320

The three-statement model's FCFF equals the driver-DCF's FCFF to the rupee, and the value is ₹320: one set of drivers, two presentations. The full model earned its keep by producing the balance sheet and the checks.

5. Extend: the receivables stress

stress = project(hist, Assumptions.kaveri_base(nwc_pct=0.277))     # NWC stays at FY26's 27.7% of revenue
fcff_s = fcff_from_projection(stress)
tv_s = terminal_value(stress["income"].loc["nopat"].iloc[-1] * 1.055, w, 0.055, ronic=0.18)
ps_s = equity_bridge(dcf_fcff(fcff_s.tolist(), w, tv_s)["ev"], 140.5, 18.2) / 6.07
print(round(ps_s), round(stress["balance"].loc["cash", "FY27"], 1))
# ₹297 (−7% vs base); FY27 cash 48.0 instead of 87.1

Notice what the stress does not do: it does not write anything off, cut growth or margins. It only says "receivables never normalise", and that alone removes ₹23/share and ₹39 Cr of FY27 cash. Add the write-off and the growth cut and you are at the reference bear case (₹168). Try:

margin_down = project(hist, Assumptions.kaveri_base(ebitda_margin=[0.133, 0.138] + [0.14] * 8))   # → ₹277
revolver = project(hist, Assumptions.kaveri_base(nwc_pct=0.30, min_cash=40.0))                    # watch revolver_draw in cf

min_cash switches on the revolver: if closing cash would fall below ₹40 Cr, current borrowings are drawn to cover it (cf.loc["revolver_draw"]), and interest follows the next year. That is how a model answers "can the company fund this scenario?" rather than merely "what is it worth?".

6. Point it at a real company

from fi.data import fetch_statements, fetch_price_info, revenue_row
t = "KIRLOSBROS.NS"                     # a real pump maker — any NSE ticker; disable VPN first
stm = fetch_statements(t, period="annual", in_crore=True)
inc_r, bs_r, cf_r = stm["income"], stm["balance"], stm["cash"]
print(inc_r.index.tolist()); print(bs_r.index.tolist())      # inspect row names — they vary

Then build a history frame in the Kaveri layout — the rows project() needs are listed in its docstring: rev, mat, pat, net_block, cwip, rou, intangibles, inventory, receivables, oca, cur_inv, cash, share_capital, other_equity, lt_debt, st_debt, lease_liab, payables, ocl, dtl. Map them from the Yahoo labels (e.g. total_revenue → rev; cost_of_revenue → mat; net_ppe → net_block; inventory; accounts_receivable; cash_and_cash_equivalents; long_term_debt/current_debt; accounts_payable; stockholders_equity less common_stock → other_equity). Fill missing rows with zero and say so. Then:

import pandas as pd
history = pd.DataFrame({...}, index=[...])          # your mapped rows × years, last column = latest FY
a = Assumptions(growth=[0.10] * 10, ebitda_margin=0.12, nwc_pct=0.20, lt_debt_change=0.0)
p = project(history, a); print(p["checks"].round(3))

The checks will pass (the model balances by construction) — which is exactly why the historical tie-out in 10.2 must be done first: project() starts from your last column and cannot know whether it was right. Then set drivers from the company's own history (fi.ratios.ratio_dashboard on your mapped frame), value it, run the reverse DCF at the market price, and write down what the price implies.

7. What the model simplifies

Simplification (model.py)	Effect	When to extend
Lease liabilities held flat; new leases treated as part of capex	FCFF unaffected; BS lease line static	Lease-heavy companies (retail, QSR): model ROU additions, amortisation, payments as generate.py does
Interest on opening balances	Slight interest understatement in draw years	Highly levered or fast-changing debt: iterate
Dividends = payout × prior-year PAT	Timing convention	Interim dividends; buybacks: add explicitly
Tax fully cash; DTL flat	Cash tax ≈ book tax	Companies with large tax-depreciation timing differences
Single entity, no NCI, no associates	Bridge is net debt + leases only	Groups: add NCI and investments to the bridge and to the model
No exceptional items in forecasts		Add a row when a known one-off is coming (a land sale, a write-off)
Margin as an input path	No cost build	When margin is the thesis: build materials/employees/other separately

Every one of these is a conscious choice you can see in the code; that is the advantage over a spreadsheet you inherited.

Trader's lens

You have now built the equity analyst's pricer: parameters (drivers) → cash-flow surface (statements) → price (value), with risk outputs (scenarios, stress, checks). The next step in your world would be to calibrate it to the market (reverse DCF) and trade the difference between your parameters and the implied ones. That is Module 11.

Common mistakes

• Running project() on unreconciled historicals (it will balance anyway — and be wrong).
• Forgetting in_crore=True and mixing rupees and crore.
• Trusting Yahoo's row labels without printing them.
• Changing drivers without recording why (keep a dated assumptions log).
• Reading the value and ignoring the balance sheet the model produced.

Key terms

Term	Meaning
Assumptions	The fi.model dataclass holding every forecast driver
project()	The fi.model function that turns history + assumptions into three balanced statements and checks
fcff_from_projection()	FCFF series derived from the projected statements
Revolver (min_cash)	Model logic that draws short-term debt to keep cash above a floor
Row mapping	Translating a data source's labels into the model's required rows
Reference valuation	The course's base case (₹320) that the build must reproduce

Check your understanding

1. Reproduce ₹320 and then change only the dividend payout to 60%. Does the value change? Should it?

Answer

Value per share is unchanged: FCFF is pre-financing and the bridge uses FY26 net debt. The balance sheet changes (less cash accumulates). In reality a higher payout could change value only via what the cash would otherwise have earned — which the model captures as other income, excluded from FCFF by design.

1. Set nwc_pct=0.30 and min_cash=40. In which year does the revolver draw, and what does that tell you?

Answer

Run it: with NWC at 30% of a growing revenue base, working capital absorbs more than operating cash in the early years and the revolver draws (check `cf.loc["revolver_draw"]`), raising interest the following year. It tells you the company's working-capital lines, not its term loans, are the funding risk — consistent with [04.5](../04-financial-analysis/05-leverage-solvency-liquidity.md).

1. Why does the model's FY27 PAT (₹96.1 Cr) exceed the "adjusted" FY26 PAT by only ~6% when EBITDA grows 6%?

Answer

Because finance costs *rise* in FY27 (interest on the higher opening debt of FY26; ₹18.3 vs ₹17.0) and other income *falls* (lower opening cash), offsetting part of the EBITDA growth; from FY28 both reverse and PAT grows faster than EBITDA. The linked model shows the financing drag the DCF does not.

1. You map a real company and project() shows cash going negative in FY27 with no revolver. What are the two possible causes?

Answer

Either the drivers genuinely imply a funding gap (high NWC or capex vs cash generation — set `min_cash` or model a debt draw), or the mapping is wrong (e.g., current debt maturing that the company will refinance was treated as a repayment; a missing row set to zero). Check the historicals tie first.

Go deeper

• tools/examples/build_kaveri_model.py — the complete script; tools/tests/test_model.py — the tests that guarantee it balances and reproduces ₹320.
• tools/running_example/generate.py — a fuller schedule (leases, CWIP, exceptional items) worth reading as a template for detailed models.
• 15.3 Worked capstone — the same model inside a full case study.

---

← Previous: 10.4 Scenarios, sensitivities & model QA · Module index · Next: 11.1 Idea generation →