An oversold reading means different things in different trends. Above the 200-day average it is a pullback that long-term buyers absorb; below it, the same reading can be the start of a real decline. Connors' RSI rules use exactly this 200-day filter to avoid catching falling knives. Applied to the deployed RSI-bounce signal, it should keep the good rebounds and drop the bad ones.
short-term mean reversion / trend filter
RSI bounce with 200-day trend filter
The idea
How it was tested
Pre-registered on 2026-09-16, with one binary change and no threshold grid. The base signal stayed as deployed: RSI(14) below 35, price within 10% of the 50-day average, at least $3M of daily dollar volume. The only addition: price above its 200-day average. Buy at the next open, hold 5 sessions, 2×ATR stop, 2.5×ATR target, most oversold first. Sizing was declared before the run at 13 slots, and the test used the current 3-month regime window.
Where it died
Gate 7 came out at t −0.01, −0.003% per trade over 271 trades. Of the 163 trades a portfolio could take, the captured alpha had t −0.32, so Gate 8 failed too. That Gate 8 figure came from the simulator before its 2026-09-19 slot-leak fix, which does not matter once Gate 7 has failed.
It was the fourth pre-registered variation of the RSI-bounce signal in two weeks, after the holding-period study, the stress gate and the RSI tightening, and like them it did not improve on the base.
The record
Thesis
An oversold RSI reading inside a long-term uptrend (price above its 200-day average) is a pullback that structural buyers absorb; below the 200-day average the same reading can be the start of a real downtrend. Adding Connors' trend filter to the deployed RSI-bounce signal should keep the good rebounds and drop the falling knives.
Specification
- signal
- RSI(14) < 35; within 10% of the 50-day MA; 20d dollar volume > $3M; price above its 200-day MA (the only change)
- entry
- next open
- exit
- 5 sessions, 2×ATR stop, 2.5×ATR target
- ranking
- most oversold first
- sizing
- declared before the run — 13 slots
- window
- current 3-month regime window
Gates
Pre-registrationpass
Was the hypothesis, its parameters and its pass/fail rule written down before any data was looked at?
Gate 7 — Jensen alphafail
After regressing out market exposure, is there alpha left?
t ≥ 1.65
t (HAC)-0.01alpha / trade %-0.00trades271
Gate 8 — Capture-preserved alphafail
Does the subset of trades a real portfolio can actually take keep the alpha?
t_cap ≥ 1.65 and ratio ≥ 0.50
t captured-0.32capture ratio+0.00trades captured163
Measured before the 2026-09-19 portfolio-simulator slot-leak fix; moot, since Gate 7 already fails.
Why it died
The trend filter did not isolate the good rebounds — it removed them. Alpha per trade fell to about zero (−0.003%) on 271 trades, in the same window where the unfiltered signal still showed positive alpha.
What it taught
A rule documented for trending markets did not transfer to a rotating one. One unverified reading is that the rebounds that paid were in names below their 200-day average; that inverted filter was recorded as a possible new hypothesis, not tried as a retry. The unfiltered signal this was meant to improve was itself archived a month later.
Verdict history
- 2026-09-16Pre-registered: one binary filter, no threshold grid, sizing declared
- 2026-09-16Single run: Gate 7 t −0.01, Gate 8 t_captured −0.32 No alpha