● alert backtest — would arming this have been worth it? No look-ahead, explained

Before you arm an alert, measure its historical edge.

Pick a price/indicator recipe — "RSI crosses below 35", "price crosses up its 20-day average" — and AtlasVector replays it bar-by-bar over history. A "fire" is counted once when the condition flips false→true (just like a live alert de-dupes), then we measure what price actually did next over 1 / 5 / 21 bars: hit rate, average forward move, and how often it false-signals. No look-ahead. The replay runs on real price history only.

NVDA⚑ RSI(14) falls below 35 (oversold)

⚑ The recipe: RSI(14) falls below 35 (oversold) — the 14-period RSI dips below 35, meaning the stock is getting oversold and may be due for a bounce. expects price to rise

HAS EDGE · triggered 2× over 260 bars

5-bar hit rate 100%avg forward move +8.30%median +8.30%edge vs baseline +7.71%head-fake rate 0%

Fires2

Hit rate100%

Avg move (5b)+8.30%

Head-fakes0%

What happened after each fire, by horizon

HorizonResolved firesHit rateAvg moveMedianHead-fake

1-bar2100%+6.73%+6.73%0%

5-bar · headline2100%+8.30%+8.30%0%

21-bar2100%+16.48%+16.48%0%

A fire at bar t is only scored at horizon h if bar t+h exists — outcomes we can't yet observe are dropped (censored), never guessed. Hit rate near 50% is a coin flip; meaningfully above (for the direction the recipe wants) is the start of a real edge.

This bullish alert fired 2×; over 5 bars it went your way 100% of the time (avg +8.30%), head-faking 0% of the time.

How often does it fire?

Your alert fired 2 times over 260 bars.

We replay your exact condition bar by bar. A "fire" is counted once at the moment the condition flips from false to true — the same way a live alert de-dupes so a condition that stays true for ten bars pings you once, not ten times.

→ Few fires — treat the stats as a rough read, not gospel.

Did the move go your way? (5-bar hit rate)

Over the next 5 bars after each fire, price moved in the expected (up) direction 100% of the time, averaging +8.30% (median +8.30%).

Hit rate is the share of fires where the forward return's SIGN matched what the trigger implies — a "bullish" alert wants price to rise. A coin flip is ~50%; meaningfully above (for the direction you want) is the start of a real edge. Mean can be skewed by one big move, so we show the median too.

→ Above-chance — promising edge worth arming.

How often does it head-fake? (false-signal rate)

0% of fires went against you at some point within 5 bars before the horizon was up.

False-signal rate measures the "it triggered, then immediately reversed" experience — the share of fires where, at any bar inside the horizon, price moved adversely to your intent. A high number means lots of whipsaw even when the endpoint eventually works.

→ Reasonably clean follow-through.

Glossary

Fire: One moment the alert condition flips from false to true; counted once per flip, like a de-duped live alert.
Hit rate: Share of fires whose forward-return sign matched the alert's directional intent. ~50% is a coin flip.
Forward return: close[t+h] / close[t] − 1 — the percentage price change over the next h bars after a fire.
False-signal rate: Share of fires that moved adversely to your intent at any bar within the horizon (whipsaw).
Horizon: How many bars forward we measure the outcome (here 1, 5, and 21 bars).