USDJPY journal: setup tags that exposed a dead edge
I kept taking the same USDJPY “pullback after impulse” scalp because it felt familiar. Expectancy looked soft but not broken — until I forced setup tags and filtered by tag for a full month of London and Tokyo samples.
The tag told an ugly story: the variant I took into thin Tokyo hours was noise after costs; the same structure after a clean London impulse still paid. Without the tag, I was averaging a dead slice with a live one and calling the blend “my USDJPY edge.” That is how zombie setups survive.
Desk change that followed:
Every entry gets one setup tag, one session tag, one cost note in R.
Weekly review is by tag, not by pair headline P&L.
Any tag under a minimum sample with negative expectancy after costs gets benched, not “tweaked forever.”
Tags are not bureaucracy. They are how you fire a zombie setup without a long argument with yourself.
Which journal tag has killed a setup you were emotionally attached to?
On the "minimum sample" before benching: it helps to know how little a small sample can tell you. If individual trades vary by roughly 1R, the standard error of the average after n trades is about 1/√n R. At 30 trades that's around 0.18R, so an average of −0.05R and one of +0.2R are hard to tell apart. At 100 trades it's 0.1R.
So I use a small sample only to bench a tag that's clearly bad, something like −0.3R or worse after 30 trades. Anything close to zero keeps running at reduced size until it reaches about 100. Your Tokyo slice sounds like the clear case.
PropScalpDesk's standard error point gets sharper once you count how many slices the tags create. With four setup tags and three session tags there are twelve cells, and a month of maybe 120 USDJPY trades leaves about ten trades per cell.
At ten trades the error on the average is roughly 0.32R, so almost nothing can be benched on one month. LondonScalper's Tokyo slice probably stood out because it was both bad and large, which is the only kind of case a single month can settle.
I keep the session tag but review setups on a rolling three months. That gives most cells 30 or more trades before I trust a verdict either way.
The tag that killed one of my favourites was a cost note rather than a setup tag. My USDJPY breakout entries in the first minutes after the Tokyo open looked flat on gross pips, but adding the spread at entry showed they lost about 0.15R a trade on cost alone.
I liked that setup because it was fast and felt clever. Do you net costs per trade in the journal, or only at the monthly level?
To answer my own cost question from last time, I now net it per trade. The journal row stores the spread at entry and the commission as a fraction of R, so a tag's expectancy is net from the first trade rather than adjusted once a month.
The other leak I found was when the tag gets written. For a while I tagged trades after they closed, and when I compared that with a month of tagging before the click, about one loser in six had drifted into a weaker sounding tag.
That drift flatters the setup you like, because its bad trades quietly move into the bin you already distrust. So the tag now goes in with the order, picked from a hotkey list of four, and it cannot be edited afterwards.
Benching also needs a way back, or a tag sits in limbo forever. Mine returns only after 40 trades on demo inside the session it failed in, judged on the same net figure.
LondonScalper's Tokyo variant is a fair test of that rule. Demo fills are kinder than live ones in thin hours, so a demo run might never show why it died, which makes me wonder whether you benched it outright or kept it at small live size?