Let's break down exactly what's happening here—this is a common gotcha with TA-Lib's candlestick pattern functions.

The Core Issue: TA-Lib needs historical context for pattern recognition #

TA-Lib's CDLHAMMER doesn't just look at a single candlestick in isolation. Hammer patterns are reversal signals, which means TA-Lib's implementation requires context from the previous candlestick to confirm the pattern. Specifically, it checks things like:

• Whether the current candle is forming after a downward move (comparing to the prior candle's close)
• The relative size of the candle's body vs. its lower wick

When you set skiprows=5 and only read 15 rows, your target candle (2020-10-23 13:15:00) ends up being one of the first few rows in your DataFrame. TA-Lib doesn't have access to the prior candles that came before your skipped rows, so it can't verify the downward trend prerequisite for a hammer. Without that context, it returns 0 because the pattern's full conditions aren't met.

In contrast, when skiprows=1-4, your target candle is positioned far enough into the 15-row dataset that TA-Lib can access the immediately preceding candle in your loaded data, allowing it to correctly evaluate all the hammer's criteria.

How to Fix This #

To avoid this inconsistency, you need to ensure TA-Lib has the necessary historical context:

• Include extra leading rows when slicing data: If you're targeting a specific candle, load at least one extra row before it so TA-Lib can reference the prior period. For example, if your target is in row 10 of your sliced data, make sure rows 9 (and maybe earlier) are included.
• Load the full dataset first, then filter: Instead of using skiprows and nrows to slice during loading, read the entire CSV into a DataFrame first, then subset to your target date range. This way, TA-Lib can access all the prior candles it needs for pattern checks.

Example Fix #

Instead of slicing during read:

# Problematic approach
df = pd.read_csv('RTS-MINUTES15.csv', header=None, names=['index', 'OPEN', 'HIGH', 'LOW', 'CLOSE', 'VOL'], index_col=0, skiprows = 5, nrows = 15)

Try loading full data first, then filtering:

# Better approach
full_df = pd.read_csv('RTS-MINUTES15.csv', header=None, names=['index', 'OPEN', 'HIGH', 'LOW', 'CLOSE', 'VOL'], index_col=0)
# Subset to your target 15 rows, plus at least 1 extra leading row
target_df = full_df.loc['2020-10-23 12:00:00':'2020-10-23 15:45:00']
target_df['HAMMER'] = talib.CDLHAMMER(target_df['OPEN'], target_df['HIGH'], target_df['LOW'], target_df['CLOSE'])

内容的提问来源于stack exchange，提问作者Sultan Dadakhanov