Study 6: Does Retail Participation Change Which Indicators Work?
We tested whether Bollinger Bands outperform RSI/KC on retail-heavy stocks. The BB lower hypothesis didn't confirm — but three sharper findings emerged: crypto-adjacent stocks broke every signal, NVDA responded to BB not RSI, and institutional stocks showed the strongest overbought reversal edge.
The Theory
Every technical indicator is, in some sense, a self-fulfilling prophecy. If enough traders watch the same level and act on it, the level works — not because of any mathematical property, but because of collective behavior.
Bollinger Bands are especially visible. The lower band (SMA-20 minus 2 standard deviations) is printed on millions of retail trading screens. The theory we set out to test: BB signals should be stronger on retail-heavy stocks because when enough retail participants watch the same chart, bounces from the lower band become self-reinforcing.
The corollary: RSI and Keltner Channel signals might be more robust across all asset types because they measure momentum exhaustion and volatility displacement — properties that hold regardless of who’s trading.
We tested it across 15 stocks split into three groups by ownership profile.
Study Design
Universe: 15 stocks across three institutional ownership buckets.
| Group | Tickers | Ownership proxy |
|---|---|---|
| Institutional | COST, JNJ, PG, KO, MMM | >70% institutional |
| Mixed | NVDA, NFLX, AAPL, BAC, MSFT | 50-70% |
| Retail | GME, PLTR, MARA, COIN, SOFI | <50%, high retail flow |
Signals tested: RSI<30, KC lower touch, BB lower touch, BB upper touch (overbought reversal), OBV below SMA, RSI+KC confluence, BB+RSI confluence.
Method: 5-year daily data (2020-2025), 10-day forward returns, 5,000-permutation bootstrap, Bonferroni correction (α = 0.05/105 = 0.000476). GME squeeze period (Jan-Mar 2021) excluded. Returns capped at ±20% per window to prevent outlier distortion.
What We Expected vs. What We Found
Expected: BB lower Cohen’s d would be meaningfully higher on retail stocks than institutional.
Actual:
| Signal | Institutional | Mixed | Retail | Retail advantage |
|---|---|---|---|---|
| BB lower | +0.046 | +0.144 | -0.028 | -0.074 |
| BB upper (reversal) | -0.286 | -0.324 | -0.117 | +0.169 |
| RSI+KC | +0.028 | +0.010 | +0.213 | +0.184 |
| KC lower | +0.061 | +0.186 | +0.159 | +0.098 |
The BB lower hypothesis didn’t confirm — retail stocks didn’t show a stronger bounce from the lower band. But three sharper findings emerged.
Finding 1: Institutional Stocks Respect the Upper Band
The clearest signal in the entire study was BB upper as an overbought reversal on institutional stocks.
When COST, JNJ, PG, KO, or MMM touches the BB upper band, the subsequent 10-day return is negative (d = -0.286 averaged across the group). The stock tends to revert back inside the band. Mixed stocks show the same pattern even more strongly (d = -0.324, led by NVDA at d = -0.617).
Retail stocks barely revert at all (d = -0.117). When GME or PLTR touches the BB upper band, momentum often continues — the band is not a ceiling.
Why this makes sense: Institutional managers run against a benchmark. When a position becomes extended and hits an overbought threshold, systematic rebalancing and profit-taking kick in. Retail investors don’t have that constraint — they hold winners.
This gives us a new tradeable signal: BB upper touch as a short-side entry on large-cap institutional stocks, independent of RSI. Not tested in prior studies.
Finding 2: NVDA Responds to BB Lower, Not RSI
In Study 2, NVDA RSI<30 produced Cohen’s d = -0.46 — a strong signal inversion. We concluded “growth stocks don’t mean-revert.”
Study 6 adds nuance: NVDA BB lower touch produced d = +0.428 — the strongest single-ticker result in the entire study.
What’s different? Bollinger Bands expand with volatility. NVDA’s 20-day standard deviation is so large that when it actually touches the lower band, it has experienced an extreme displacement relative to its own volatility regime. RSI<30 catches a different condition on NVDA — it can fire during a multi-week momentum breakdown before the lower band is even reached.
In other words: BB lower is a more extreme filter on high-volatility stocks than RSI<30. When NVDA touches its own lower band, the overshoot is genuine and tends to revert. When its RSI hits 30, the stock may still be in a structural downtrend.
This refines the rule from “don’t use mean-reversion signals on NVDA” to “use BB lower, not RSI, on high-vol growth stocks.”
Finding 3: Crypto-Adjacent Stocks Are Their Own Category
MARA and COIN broke every signal tested:
| Signal | MARA d | COIN d |
|---|---|---|
| RSI oversold | -0.261 | +0.095 |
| BB lower | -0.195 | -0.160 |
| KC lower | -0.075 | +0.137 |
| RSI+KC | -0.186 | +0.266 |
| OBV weak | +0.028 | -0.357 |
MARA in particular shows consistently negative d across mean-reversion signals — the same inversion pattern as NVDA in Study 2, but worse. These aren’t just growth stocks with momentum; they’re assets that trend for months on macro crypto cycles and then reverse sharply. There is no “fundamental anchor” to revert to.
Crypto-adjacent stocks (mining companies, crypto infrastructure, leveraged crypto ETFs) require a completely different framework. The same indicators that work on SPY with d = +0.30 produce d = -0.26 on MARA. Trading them with standard mean-reversion signals destroys capital.
COIN shows somewhat more mixed results (RSI+KC d = +0.266) — as a business with real revenue, it sits closer to the high-vol growth category than MARA.
The Surprising Retail Finding: RSI+KC Works Best
The biggest surprise in the data: RSI+KC showed the highest retail advantage of any signal tested (+0.184).
GME post-squeeze RSI+KC: d = +0.500 (the highest single result in the study).
SOFI RSI+KC: d = +0.354.
COIN RSI+KC: d = +0.266.
Why? These “retail” stocks — post-squeeze GME, SOFI, PLTR — actually have a fundamental anchor. GME post-squeeze is a real business trading at a real multiple. SOFI is a fintech with deposits and loans. When they become oversold (RSI<30) and displaced (KC lower), they revert just like defensive stocks — often more aggressively because the volatility creates a larger displacement.
The retail/institutional split wasn’t the right axis. The right axis is: does this asset have a fundamental anchor it can revert to?
The Four Behavioral Archetypes
| Archetype | Examples | Best signal | Why |
|---|---|---|---|
| Fundamental reverter | SPY, WMT, JNJ, SOFI, GME post-squeeze | RSI+KC, RSI≥50 exit | Stable business; fear-driven overshoots revert |
| High-vol growth overshooter | NVDA, NFLX, AAPL | BB lower (not RSI) | Volatility so high that BB lower = true extreme |
| Institutional overbought reverter | COST, JNJ, PG, KO, MMM | BB upper short | Systematic rebalancing creates ceiling |
| Crypto-adjacent | MARA, RIOT, BITF, leveraged crypto ETFs | None | No fundamental anchor; mean-reversion signals invert |
No Bonferroni survivors were recorded in this study — the 105-test correction at α = 0.000476 is strict. The results reflect directional signal, not certainty. Larger sample sizes and longer time windows are needed before trading these findings at full size.
What Changes in Practice
Use BB lower on NVDA/NFLX instead of RSI<30. Previous guidance was “don’t trade these.” Refined: RSI<30 doesn’t work, BB lower does — especially when it coincides with a VIX spike.
Add BB upper as a standalone signal on institutional large-caps. When COST, JNJ, PG, or similar touches the BB upper band, d = -0.286 to -0.437 on the reversal. This is a new signal not in the prior toolkit.
Never apply mean-reversion signals to MARA, RIOT, BITF, or similar crypto-mining stocks. The signal inverts. Category: no-trade zone for this framework.
Post-correction retail stocks (GME, SOFI) respond to RSI+KC just like defensives. The “retail” label is misleading — they have real businesses and revert normally once the speculative premium is reset.
Next Questions
- Does BB lower on NVDA survive a dedicated Bonferroni test with a larger sample?
- Is BB upper on institutional stocks a tradeable short, or does it require additional confirmation (VIX level, RSI reading)?
- Where exactly does the crypto-adjacent category begin and end? (COIN sits in a gray zone.)
- Can we build a behavioral profiling tool that auto-classifies new tickers into these four archetypes before trading them?
The last question points toward a systematic profiling layer — a tool that measures each new stock’s mean-reversion coefficient, volatility regime, and crypto/macro correlation before routing it to the right indicator framework. That’s the next build.