The Takeover Backtest: 1,059 Hits, No Reliable Way to Predict Them

Spotting a takeover target before the buyer officially knocks is one of the oldest dreams in the stock market: buy cheap, hold briefly, and when the announcement lands, the price jumps double digits. This study puts that idea on two legs. First the count: how often does the pattern of a cash takeover even show up in the price history, and what does it look like? Then the real test: can the target be identified BEFORE the announcement from measurable company traits — and does it make money? This study answers the first question with a yes, and the second with a clear, and we think instructive, no.

Part 1: What we counted first

We searched every delisted US price series for the signature of a cash takeover: a price jump on the announcement day, followed by a tight sideways range through the last trading day. The headline number is the primary variant — a jump of at least 15%, a range of at most 5% — measured on adjusted closing prices. We checked 11,144 of 11,144 delisted US stocks with a price series — a complete run, not a sample. One framing point matters here: the signature is identified LOOKING BACKWARD, on a price series that has already ended. Part 1 is a stocktake, not a trading idea — whether the same pattern can be spotted BEFORE the announcement is exactly what Part 2 tests.

Setup: survivorship-free, twelve threshold variants

So the result does not just reflect the "survivors," only stocks that have since been delisted are included — the delisting itself is part of the signature. The jump is measured close-to-close, because the price source carries no opening prices; since announcements arrive overnight, that is functionally equivalent. To show how much the result depends on the exact threshold definition, the same search was repeated across twelve combinations of jump size (10/15/20/25%) and maximum range (3/5/8%).

The result of the count

The tight range lasts a median of 49 trading days (Q1: 30, Q3: 67) until delisting; the final price sits just 1.0% above the jump level at the median. That picture — a jump, then a nearly motionless range until the price series ends — is exactly the signature Part 2 tries to identify in advance.

How sensitive is the count to the thresholds?

The count reacts to the threshold as expected but stays within the same order of magnitude — no combination delivers a multiple of the primary variant. About 27.8% of hits also carry the "artificially extended price series" flag (trailing filler bars were trimmed off), 7.0% trade under $1, 3.3% are flagged as a possible unadjusted reverse split, and 3.0% show an implausible premium — these cases stay in the count but can be individually excluded via the flags.

Cross-check against the SEC

Of 37 checkable hits in a calibration sample, the SEC (EDGAR) confirms 37 (100.0%) with a genuine takeover filing — documented among others by 25-NSE, DEFM14A, PREM14A, 15-12B, and SC 14D9 forms. False positives from stock swaps or bankruptcies: 0 of 14. At the same time, the same calibration sample shows a flip side: of 37 known cash deals, the primary variant recovers only 12 (32.4%) — names like ATVI, CERN, CTXS, RHT, SPLK, or TWTR are missing, mostly because the deal took more than 380 trading days to close or the announcement leaked in advance, shrinking the jump. The counted figures are therefore a floor, not a complete count.

Part 2: Can the target be spotted in advance?

Part 1 counted how often the pattern occurs. Part 2 asks whether the target can be identified from company traits BEFORE the announcement — and whether that makes money. Those are two separate questions: the first could, in principle, answer yes while the second still answers no.

The universe is a monthly grid from January 2012 through July 2025: entry requires a raw closing price of at least $1.00 and a market cap of at least $100 million, with an annual report already published as of the cutoff date (no more than 18 months old). The label is the Part 1 jump in the loosest variant (10% jump / 8% range) within twelve months AFTER the cutoff date, excluding hits with artifact flags.

The point model: twelve traits, transparently weighted

Rather than a black-box trained classifier, this is a simple, open point model: each of the twelve traits carries a direction and weight fixed in advance, and a stock only receives a score if at least 60% of the total weight is measurable. The rank is built within the cross-section of the SAME month, not against absolute thresholds.

The single largest weight goes to the "growth mismatch": cheap, resource-rich, AND growth-starved at the same time — the core hypothesis of arguably the most influential academic work on this subject, Krishna Palepu's 1986 study. Palepu himself, writing with considerably more caution than modern marketing copy, put it this way:

"Several published studies claim that acquisition targets can be accurately predicted by models using public data. […] The results show that it is difficult to predict targets, indicating that the prediction accuracies reported by the earlier studies are overstated."

— Krishna G. Palepu, "Predicting Takeover Targets: A Methodological and Empirical Analysis," Journal of Accounting and Economics, 1986.

Nearly forty years later, with a far larger dataset and far more computing power, this study confirms Palepu's core finding — even though the tools today are different.

No number from the future: the walk-forward proof

An application year's score threshold is calibrated only from panel months through December two years earlier — not through the end of the prior year, because a month's label is not known until twelve months later. In addition, only hits whose last trading day falls before January 1 of the application year are counted: a deal is only classifiable as a takeover once its price series actually ends, and that lag runs past the calendar year boundary for close to a quarter of all deals. The first application year is 2017.

All three checks are clean — every required number comes exclusively from before the respective application year.

Is the hit rate elevated in the top decile?

The following table is the honest one: the decile comes from a score threshold calibrated exclusively from earlier years. The deciles come out unequal in size (38,885 to 46,611 stock-months): because the score thresholds come from earlier years, they split the later cross-section into only approximate tenths, not exact ones. And the base rate here reads 0.94% rather than the panel-wide 0.83%, because the lift table only covers the application years from 2017 onward — the walk-forward test needs earlier calibration years, and within the application years themselves the base rate runs higher than the panel average.

No single decile stands out clearly, and the order is not monotonic: decile 1, theoretically the strongest, sits below chance at a 0.92 lift; decile 4 leads instead at a 1.16 lift — a pattern that looks more like noise than signal. For comparison, the same calculation using the within-month rank — which needs no calibration and is the ceiling of what the model could achieve — puts the top decile at a 0.88 lift, also below the base rate.

Top decile by year

Year by year, the top decile's lift also swings without a discernible pattern, between 0.27 (2019) and 1.42 (2025) — exactly what you would expect from a model with no real predictive power.

Does it make money? The portfolio simulation

Raw hit rates say nothing about what a portfolio with limited capital would have actually earned. We simulated: a monthly purchase of the top decile, a twelve-month holding period, 0.1% cost per leg (buy and sell), forced sale on delisting.

The portfolio loses to both benchmarks — not just to the index, but also to its own equal-weighted investment universe, which is subject to the same delisting treatment as the portfolio and is therefore the fairer comparison.

Only 57 of 5,917 positions purchased (0.96%) were actually acquired within twelve months. The median return per position of 1.27% — far below the mean of 10.67% — shows the same pattern seen elsewhere: a small number of outliers carries the average, while the typical position is barely more than a wash.

How good is the label, really?

A model can only learn what the label contains. Of 37 known cash takeovers in the reference list, the label variant used for calibration (10% / 8%) recovers 19 — a recall of 51.35%. The SEC cross-check stays spotless: 37 of 37 checked hits are confirmed by a genuine takeover filing. The measured lift is therefore the conservative figure, not the flattering one — real takeovers the detector misses count as zero in the label even though they should be a one.

Limitations of this study

• Only the cash-offer signature. Only the pattern of a cash takeover is learned and counted. A stock-swap deal never shows it — the price keeps tracking the acquirer's stock right up to the end and falls entirely outside the label. The measured hit rate therefore understates the true takeover frequency.
• The comparison index is a pure price index. The S&P 500 is available only without dividends. A portfolio that collects distributions would carry a structural edge of roughly two percentage points a year against a price index — the more meaningful benchmark is therefore the equal-weighted universe, treated identically to the portfolio. And the portfolio loses to that one too.
• The panel covers only larger, liquid stocks. Only stocks priced above $1 with a market cap above $100 million are included. This study makes no claim about takeovers of smaller companies.
• Trading costs likely understate reality. The model assumes 0.1% per leg; bid-ask spread, market impact, and taxes are NOT included — for smaller, less liquid names the real drag would be noticeably higher.
• The Part 1 detector does not find every known cash takeover. The recall of the label variant used for Part 2 sits at 51.35% against the reference list — everything beyond that counts as zero in the label even though it should be a one. The measured lift is therefore the conservative figure.
• A supplementary logit regression check was skipped. The headline result of this study was always meant to be the transparent, fully auditable point model — the logit would only have confirmed its direction, not added a new finding.

Readers who want to see the same survivorship-free rigor applied to a different pattern can find it in our Qullamaggie Backtest — another backtest study in this series. Readers looking for warning signals BEFORE a bankruptcy rather than BEFORE a takeover will find that in our Insolvency Radar Top 10 study. More backtest studies live together in Studies, and our continuously updated pattern scanners live under Stock Scanners.

Figures as of August 7, 2026. Price data through August 3, 2026.

This article is a historical analysis of publicly available price and fundamental data and not investment advice. It contains no buy or sell recommendation, no forecast, and no statement about any individual company listed today. Past results — whether simulated or real — are not a reliable indicator of future returns. Anyone making investment decisions should assess their own situation and risks, if in doubt with professional advice.