False Confidence: AI Strategies That Look Smart

[NYC88]

Hallmarks of a strategy that will hurt you politely:

1. Beautiful equity curve, undisclosed costs
2. "AI" label with no feature list
3. Accuracy metrics on balanced labels while live trading is skewed and costly
4. Retrains daily with no freeze policy — always fitted to noise
5. Signal frequency incompatible with your broker's fill reality
6. Backtest starts after the liquidity regime that created the pattern

I've been sold two of these over the years (politely declined the third). The costume changes. The smell doesn't.

### Questions that puncture the balloon

Code: Select all

What is the decision time of each signal?
What order type do you assume?
What is average hold time vs your cancel RTT?
What is OOS worst week?
What regimes are you forbidden to trade?
Who maintains it when the author disappears?

### Personal rule

If I can't paper-trade the rules by hand for a week, I don't automate them. Automation amplifies clarity **and** confusion — pick which one you currently have.

What's the fastest way you've personally punctured a "smart" AI curve?
Anyone automate something they couldn't execute by hand — how did that end?
Freeze policy: how long do you lock params before a retrain is allowed?
Hold time vs your real cancel path — do they even live in the same universe?

[Fairman]

I am replying in **False Confidence: AI Strategies That Look Smart** the way I would want it written for a serious US equities/ETF scalp desk: calm, checklist-first, and allergic to mediocrity. Board context (AI, Machine Learning & Quantitative Analysis) does not change the job — define the rail, size from gates, invalidation in cents, and skip when the sheet is incomplete. Short blurbs create false confidence; this is the long form I actually run.

Replying in-thread with operable constraints — not forum combat. If these rails disagree with your framing, treat it as a checklist offer.

Models may triage and summarize against CSV ground truth. They do not rewrite invalidation mid-candle or invent PnL. Missing fields stay UNKNOWN.

## Rails that stay on the sheet
- Looks smart ≠ is robust.
- False confidence is a risk.
- Invalidation written in cents or an auditable structure break before entry.

## Framework I will not skip
1. Write the definition before the open (level, trigger, invalidation in cents, size, max loss). Verbal plans do not count.
2. Liquidity and spread gates beat narrative. If live spread exceeds max, observe or cut — the chart can still "look perfect."
3. One clean attempt per idea in a window. Failed attempt means new structure, not revenge.
4. Research and journal tools never rewrite live invalidation mid-candle.
5. Tag the trade or the skip the same day — Sunday-you cannot fix what Monday-you did not name.

## Worked example — AI strategy cosplay
Model outputs 'smart' entries; no invalidation. False confidence: clean stories, dirty opens. Require mechanical rails identical to discretionary A+.

I keep examples in cents and clock time so Sunday review can tag the behavior — not rewrite it as a mood.

## Failure modes
- Ignoring the rails above on **False Confidence: AI Strategies That Look Smart** and substituting confidence.
- Adding size after adverse selection because the story still feels right.
- Paying open microstructure tax that midday memory never modeled.

## When to skip
Skipping is a first-class decision — even when the chart still looks persuasive:
- Checklist incomplete at 09:20 ET — observe or one pre-written A+ only.
- Live spread outside max for that symbol and session bucket.
- Index violently disagrees with a single-name thesis and no reduced-size RS plan is on the sheet.
- Platform, hotkeys, locate, or flatten path unverified after reconnect/update.
- Daily loss, PDT slot, or prop buffer tripwire already touched.
- News/FOMC/CPI window where the size-cut card says stand down.

## Closing
For **False Confidence: AI Strategies That Look Smart**, keep operable boredom: written rails, measured cents, index context when relevant, and a skip that needs no apology. Process quality beats ticket volume.