Why No Two Market Bottoms Look Alike

Capitulation wears different clothes

The premise behind the challenge is that a good indicator should fire at every major low. Berg rejects the premise: different declines end in different kinds of data extremes. Capitulation can show up in the rate of decline into the low, in extreme downside-to-upside volume ratios, in divergences between the major indexes and the advance-decline line, in strings of consecutive downside gaps or down days, or in VIX and VXN deviations from trend.

Each bear market ends in its own way. Demanding that one indicator capture all of them is demanding that panic always take the same shape — and it doesn't.

Sometimes the bottom signals after the fact

In some cases the low itself never presents a clean, tradable oversold extreme. What triggers the buy signal instead is the momentum that emerges in the days immediately after — the character of the rebound, not the depth of the decline. Berg's analysis has modeled every major market low since 1957 and found extreme readings either going into the low or immediately following it, but the specific combination differs every time and cannot be known in advance.

The cluster is the signal

This is why the library holds hundreds of models rather than one. Individual signals are rare by design — three or four precedents each — but genuine turning points set off many of them at once. At the April 2025 bottom, signals fired on eleven separate dates between April 4 and April 30, including fifty-seven on April 9 alone; collectively the cluster corresponded to more than 45 prior market dates.

A skeptic looking at any single model sees a sample size of four. The composite sees dozens of independent rare events converging in one month. Those are very different evidentiary claims.

Regimes change — and the library knows it

Markets since 1957 have absorbed the end of the gold standard, the rise of corporate buybacks and of hedge funds, the arrival of index futures and listed options, the spread of passive and active ETFs, and a long slide in dividend payout ratios. Berg's models reflect that history: some have fired steadily across the entire 1957–2025 record, while others are silent until the early 2000s, appearing only once ETFs and program trading had reshaped the market's structure.

A model set tuned to specific regimes, retired or dormant when its regime is absent, is not a bug in the methodology. It is the methodology.

Volume regimes cut both ways

One current example of context-dependence is what Berg calls a high-volume regime — a stretch in which 5-day volume has recently hit multiyear highs. Market lows and early bull-market advances have almost always occurred inside such regimes, so when that volume arrives into or near a low after a decline, it reads as a bullish tell. When it arrives after a substantial advance instead — the S&P 500 in 1987, gold and silver in early 2026 — it has typically warned of substantial downside volatility ahead.

Same data, opposite meaning, disambiguated entirely by context. No always-on oscillator can make that distinction; a regime-aware model library can.

Reflecting boundaries

A final layer from Berg's 2008 paper “The Boundaries of Technical Analysis”: some data-driven signals identify what he calls reflecting boundaries — price zones that, historically, are never successfully revisited. At the time of the signal, nothing on a conventional chart marks these zones as support; they surface only in the underlying data. When later selling attempts to trade back into such a zone, the attempt fails and is quickly rejected.

It is one more expression of the same idea that runs through all of this work: the market's turning points leave fingerprints in the data — just never the same fingerprints twice.