I have been running a mean-reversion swing signal live for a couple months and I'd love honest critique of the logic before I trust it with more size.
The basic idea is that quality companies overshoot to the downside when the market panics. Strong business, temporary dislocation, we buy the overshoot, sell the bounce back to the mean. I only run it on a hand-picked universe of \~60 large-caps, so no random tickers and no small caps.
A buy signal has to clear **four checks**, in order and then a market-regime overlay decides sizing. The checks are where I want the feedback:
**1. Trend filter.** Price must be above its 200-day MA so we are more likely buying dips in uptrends, not catching knives in downtrends. (One pattern can fire just below it if volume strongly confirms.)
**2. Pullback zone.** It has to be 10–25% off its 52-week high. Less than 10% isn't a real dip; more than 25% on a quality name usually means there is bigger chance that something's actually broken.
**3. Oversold (RSI).** RSI has to be oversold, with the threshold scaling to the stock's volatility (ATR): ≤35 for calm names, ≤32 medium, ≤30 for jumpy ones.
I use **simple RSI, not Wilder**, and I read the daily RSI **live**, projecting the current price as today's provisional close, so I can act while the setup is forming instead of waiting for the 4pm bar to finish. Both are "catch it early before it has already bounced" calls, open to being told they dont make any sense.
**4. Smart-money confirmation.** This should act as a falling-knife guard.
I score where price closes in the day's range (Chaikin-style: close near the high = buyers won the day = positive; close near the low = sellers won = negative) **times** a volume-conviction factor (today's volume vs its 20-day high, adjusted for time of day so the open doesn't fool me).
Runs roughly −100 to +100, and I need **≥ +40** to fire — buyers demonstrably winning the range on real volume, not just allow RSI on a quiet tape. I also ignore the first 30 minutes (range is too thin early and ads noise to the score).
If all four line up, the actual trigger is one of a few **dip patterns** — a panic-volume capitulation, a fast RSI drop into the lower Bollinger band, an RSI hook turning up off it, or a volume-confirmed absorption *at* the band. (Bollinger isn't a separate gate — it's the "how stretched" measure inside these patterns: price at/below the lower band.) They all still have to pass check #4.
**Exits:** hard stop at 3×ATR, first target at the 20-day MA (the mean it reverted from), a trailing stop that arms at +10%, and a time stop if it's going nowhere after \~25 days.
**The macro overlay (mostly sizing).** Sitting on top of all this is a market-regime read — VIX level + trend, breadth, and the VIX term structure rolled into one score. Most of the time it isn't a gate at all; it just adjusts position *size*. Mean reversion actually *needs* volatility, so I size **up** in a moderately stressed tape (VIX \~20–30, where quality overshoots hardest) and **down** when things are too calm (shallow dips, weak reversion). It only becomes a hard "no new entries" switch in extremes: full-blown panic (VIX >30, where everything just gaps and keeps falling), or the 48h right after a big VIX spike (spikes cluster, so I let it settle before entering). So volatility's a feature here, not something I hide from. This overlay mostly tunes size, and only blocks entries when the tape is genuinely disorderly.
**The** **results so far:** win rate's been north of 80% over the couple months it's run, but it's been a broadly up market the whole time, and I know a high win rate for a dip-buyer in a bull is mostly beta. So I'm logging every signal's forward return vs SPY to measure actual **alpha**, not win rate. That's the question I can't answer yet.
**What I'd like feedback on, besides the general strategy:**
1. Is the smart-money score (range-close × volume-conviction) a sound confirmation, or am I amplifying noise? Any work on close-location-value as a short-term reversal predictor?
2. Simple vs Wilder RSI for *entry timing specifically* — am I fooling myself that faster is better?
3. The 30-minute ignore window — reasonable, or arbitrary curve-fitting?
4. How would *you* separate real edge from beta without a bear-market dataset? Is a couple months of an up-tape just unanswerable until I get a real drawdown?
Not selling anything, no Discord, no course. Just want to know if the logic holds or if I'm kidding myself.