Research Log — what didn’t work — Daniel Hughes

TEDGPAYING MEMBERS~2,600▸TEDGCOMMITS (SOLO)~1,800▸TEDGLINES OF CODE~120k▸GEXSYMBOLS SERVED29▸GEXSNAPSHOT CADENCE30s▸GEXLOAD-TESTED3.1k msg/s▸IBAANNUALISED RETURN+50%▸IBAMAX DRAWDOWN6.1%▸IBASHARPE2.1▸GEXRTRADES499▸GEXRWIN RATE72.7%▸GEXRPROFIT FACTOR2.60▸WHALTRADES426▸WHALRETURN+224%▸WHALSPY SAME PERIOD+57%▸BTARULES BIAS P&L$8.1k▸BTALLM BIAS P&L$2.9k▸BTALLM, DISAGREE DAYS−$271▸VOLSESSIONS72,168▸VOLETFS29▸VOLHISTORY FROM2017▸TEDGPAYING MEMBERS~2,600▸TEDGCOMMITS (SOLO)~1,800▸TEDGLINES OF CODE~120k▸GEXSYMBOLS SERVED29▸GEXSNAPSHOT CADENCE30s▸GEXLOAD-TESTED3.1k msg/s▸IBAANNUALISED RETURN+50%▸IBAMAX DRAWDOWN6.1%▸IBASHARPE2.1▸GEXRTRADES499▸GEXRWIN RATE72.7%▸GEXRPROFIT FACTOR2.60▸WHALTRADES426▸WHALRETURN+224%▸WHALSPY SAME PERIOD+57%▸BTARULES BIAS P&L$8.1k▸BTALLM BIAS P&L$2.9k▸BTALLM, DISAGREE DAYS−$271▸VOLSESSIONS72,168▸VOLETFS29▸VOLHISTORY FROM2017▸

F8 RLOG GO9 KILLED

Research Log — what didn’t work

Most ideas don’t survive contact with out-of-sample data. These are the ones I killed, and why.

Idea	Sample	Result	Verdict	What I learned
GARMA A discretionary liquidity-sweep playbook, coded rule for rule	155 days, real NQ	245 trades · PF 1.08 · negative without the top 5 wins	NO EDGE	Looked fine on a proxy price series (PF ~1.7); the edge disappeared on the real futures tape. Validate on the instrument you trade.
GARMA-Wicks Wick-and-reclaim at session levels with a volume burst	677 days, NQ 1m	1,263 trades · PF 0.90 · −77.7R total	NEGATIVE	Above zero in only 6% of bootstrap resamples. Never traded.
Mechanical gamma setups Eight rule-based setups read off intraday gamma maps	Minute-bar backtest	93 trades · PF 0.76	NEGATIVE	Map “shapes” alone don’t pay; context (time, opening range, regime) carries the edge.
LLM map reader Let Claude set the morning bias instead of rules	147 days	$2.9k vs $8.1k for rules	REJECTED	The deterministic rule wins. The LLM stays a proposer, not a decider.
IB reversal setups Sweep and flip reversals inside the IB strategy	Hold-out	Lost on the hold-out	REJECTED	Switched off in production; the core continuation setup carries the strategy.
RS Pullback Relative-strength pullbacks to weekly EMAs on the S&P 500	Two hold-outs	13,826 trades · +0.6% / trade vs random · ~0% account CAGR	NOT TRADABLE	A real but tiny edge that can’t beat holding the index after capital constraints.
Sector vol H1 Risk-reversal percentile predicts forward vol	72k sessions	Failed the pre-registered test	FAILED	Reported as a null result rather than re-specified until it passed — no p-hacking.
Sector vol H2 SMH’s surface informs XLK’s forward vol	72k sessions	Passed, then withdrawn	WITHDRAWN	Overlap placebos showed an errors-in-variables artefact from shared holdings.
Third-party risk on/off Rebuild a published daily SPY regime signal	2019 – 2026	Rebuild: 12.9% CAGR vs 32.7% claimed	REJECTED	The original was chosen from ~100k candidates in-sample; it trails SPY since 2023.

RLOG DES GO → full write-up