Candlestick patterns need context

Scan the NSE for every bullish engulfing candle and you get a long list on any day, in any market. That is the whole problem with pattern screening. A candle is two or three bars describing a short fight between buyers and sellers, and the fight means nothing until you know where it happened. An engulfing bar at a level you can name, after a decline, on volume, is worth a look. The same bar mid-range on a sleepy Wednesday is a shape.

So every scan in the pattern library here carries context conditions next to the pattern flag. Three kinds, and I would rank them in this order.

Location

A hammer is a rejection: price driven down, bought back, sellers refused. Rejection only tells you something when there was something to reject from. So hammer at oversold wants the stock genuinely stretched, RSI under 40, before the hammer counts, and bullish engulfing at support wants the engulfing bar printed within reach of the 50-day average, where there is an owner base to defend.

In Sift a pattern is a field like any other, so context is just more conditions:

pattern is bullish_engulfing within 2 bars
  and rsi(14) < 40

The within 2 bars earns its place. Patterns are events, and an event from Monday can still be fresh on Wednesday. Offset-based screeners handle that with duplicated clauses when they handle it at all.

The preceding trend

Reversal patterns need something to reverse, and the three-bar patterns carry that requirement inside their own definition. A morning star is a down bar, then a small indecision bar, then a strong up bar in that order. The flag only fires when the sequence actually printed, which is why its scan can afford to be almost bare.

Single-bar patterns have no sequence to lean on, so their scans check the trend explicitly. The shooting star only counts when it prints into strength, with RSI already above 60.

The neutral case makes the point best. A doji is pure indecision, and indecision only carries information where one side had been winning, so that scan asks for the stock to be above its 50-day average with volume running above normal before a doji is worth reading at all.

Volume

Patterns describe a fight and volume tells you how many people showed up to it. Half the pattern scans here carry a rel_volume condition for the same reason the breakout scans do.

Then make the folklore testable

Candlestick lore is old, confident, and almost entirely unexamined on Indian data. It does not have to stay that way. Every pattern scan here carries a hit-rate panel, so "does a morning star after weakness actually lead anywhere on the NSE" is a question with an answer in about ninety milliseconds: how often the scan matched across 250 sessions, and what those matches did over the next 1, 5 and 20 days, close to close, before costs.

Read it with the usual discipline and drop the patterns that fail. When I did this the survivors were a much shorter list than the textbooks suggest, which I count as the screener working. Fewer candles is the point.