I backtested my own DCF model on 13 years of point-in-time data. Cheap stocks didn't beat the market.
retained text
I built a fair value model (DCF plus normalized earnings and peer multiples, with a bear/base/bull case) and before trusting it I wanted to know one thing: if you buy what it says is undervalued, do you actually do better? So I wrote down the rules and what counts as pass/fail first, then ran it. Every quarter from end of 2012 to 2025 I valued about 450 liquid US stocks using only what was known at that date. SEC filings by the date they were filed, treasury yield of that day, industry data from that year. No analyst forecasts, growth comes only from the company's own past 3 years. Banks, REITs and loss making companies left out. Around 38k valuations total. Then each quarter I split stocks into fifths by price vs fair value and looked at the next 12 months return vs the average stock. 2013-2018: \- cheapest fifth: -3.1% vs average \- most expensive fifth: +1.6% \- basically fully backwards. the "overvalued" stocks won 2019-2025: \- cheapest fifth: +2.3% vs average \- most expensive: -0.8% \- right direction this time, but not significant (t around 0.8, so could easily be luck) If you just bought the cheapest fifth every quarter from 2019 to 2025 (equal weight, 10bp costs) you got 11.6% a year vs SPY 15.6%, and a worse drawdown (-46% vs -32%). I also checked that it's not some bug: re-ran 30 random valuations with every filing after the date deleted and got the same numbers, and the share counts match SEC cover pages. What I'm NOT saying: \- that value investing doesn't work. this is one model, one horizon (12 months), large liquid stocks only \- that DCF is useless. it's still a good way to see what the price assumes about the future Things that could be wrong with my test: \- price data only has companies still listed today, which usually makes value look better, not worse \- 12 months might just be too short for value to play out. a lot of you hold 3-5 years \- trailing growth only. with real analyst estimates it might behave differently \- sectors are today's, not historical Honestly I expected it to work at least a little. The 2013-18 part surprised me the most, cheap stocks didn't just fail to win, they lost. Curious how people here handle this. Do you think it's the horizon, the model, or that a simple "price below fair value" rule just doesn't capture what makes a good value pick (quality, moat, catalyst etc)? **EDIT (Oct 13): ran what you guys asked for. short version, nothing changed.** - **3 and 5 year holds:** cheapest fifth vs the average stock was -2.7% over 3 years and -1.6% over 5 years (whole hold, not per year). at 5 years the middle of the pack did best and the most expensive did worst, cheapest was just average. t-stats are newey-west for the overlapping windows, but 3y and 5y only have about 4 and 2 independent periods so take them as descriptive - **quality filter** (5 years of steady positive margins, low debt, revenue growing): cheap + quality was +0.4% at 12 months, -1.7% at 3y, -1.3% at 5y. and it flips by period, -3.9% in 2013-18, +4.3% in 2019-25 - **equal weight universe:** holding all ~450 stocks equal weight made 11.6% a year in 2019-25, same as the cheapest fifth. so the ranking added nothing, most of the gap to SPY is equal weight vs cap weight - **sector split:** energy and materials explain about 2 of the 3.3 points the cheap stocks lost in 2013-18 (the oil crash), but ranking within each sector it's still -2.9%. so partly the oil crash, not all of it - **base rates** (the mauboussin idea from the comments): I wrote the rules down first again and tried two versions. ranking by how unlikely the growth the price needs is (vs how often companies that size ever did it): cheapest fifth +2.0% in 2019-25 (t 0.6), -3.4% in 2013-18. and the DCF using typical growth for the company's size instead of its own last 3 years: +2.2% and -2.9%. both fail, basically the same as the original - the only thing that looked positive was quality on its own at any price, +2.9% over 3 years. but i found that after the main test, its one of about a dozen cuts, and my data only has companies still listed today which flatters quality. so a lead, not a result full numbers, the curve and every quarter are here: https://quantedge-nine.vercel.app/evidence?run=fair-value so i don't think it's the horizon or the trailing growth. either cheap-vs-fair-value just doesn't rank stocks well, or my model is missing what actually matters. thanks for all the ideas, this thread made the test a lot better
document fetched 2026-10-13 00:45 · canonical https://reddit.com/r/ValueInvesting/comments/1x3zubp/i_backtested_my_own_dcf_model_on_13_years_of