Accuracy

numta is designed to produce results that match TA-Lib's output for all implemented indicators.

Verification Methodology

numta includes comprehensive accuracy tests that compare output against TA-Lib:

# Install with comparison dependencies
pip install -e ".[dev,comparison]"

# Run accuracy tests
pytest tests/accuracy/ -v

Accuracy Guarantees

Numerical Precision

All indicators are tested to match TA-Lib output within floating-point precision:

• Relative tolerance: 1e-10
• Absolute tolerance: 1e-10

Test Coverage

Each indicator is tested with:

• Standard parameters (e.g., RSI with timeperiod=14)
• Edge cases (minimum periods, single values)
• Various data lengths (100 to 10,000 points)
• Random and trending data patterns

Accuracy Test Results

Moving Averages

Indicator	Status	Max Difference
SMA	✅ Pass	< 1e-14
EMA	✅ Pass	< 1e-14
WMA	✅ Pass	< 1e-14
DEMA	✅ Pass	< 1e-13
TEMA	✅ Pass	< 1e-13
KAMA	✅ Pass	< 1e-12
TRIMA	✅ Pass	< 1e-14
T3	✅ Pass	< 1e-12

Momentum Indicators

Indicator	Status	Max Difference
RSI	✅ Pass	< 1e-12
MACD	✅ Pass	< 1e-13
STOCH	✅ Pass	< 1e-12
ADX	✅ Pass	< 1e-11
CCI	✅ Pass	< 1e-12
MFI	✅ Pass	< 1e-12
ROC	✅ Pass	< 1e-14
MOM	✅ Pass	< 1e-14
WILLR	✅ Pass	< 1e-14

Volatility Indicators

Indicator	Status	Max Difference
ATR	✅ Pass	< 1e-12
NATR	✅ Pass	< 1e-11
TRANGE	✅ Pass	< 1e-14

Volume Indicators

Indicator	Status	Max Difference
OBV	✅ Pass	< 1e-14
AD	✅ Pass	< 1e-12
ADOSC	✅ Pass	< 1e-11

Bands and Channels

Indicator	Status	Max Difference
BBANDS	✅ Pass	< 1e-13
SAR	✅ Pass	< 1e-11

Running Accuracy Tests

Quick Accuracy Check

import numpy as np
from numta import SMA, RSI, MACD

# Try importing TA-Lib for comparison
try:
    import talib
    HAS_TALIB = True
except ImportError:
    HAS_TALIB = False
    print("TA-Lib not installed, skipping comparison")

if HAS_TALIB:
    # Generate test data
    np.random.seed(42)
    close = np.random.randn(1000).astype(np.float64) + 100

    # Compare SMA
    numta_sma = SMA(close, timeperiod=20)
    talib_sma = talib.SMA(close, timeperiod=20)
    sma_diff = np.nanmax(np.abs(numta_sma - talib_sma))
    print(f"SMA max difference: {sma_diff}")

    # Compare RSI
    numta_rsi = RSI(close, timeperiod=14)
    talib_rsi = talib.RSI(close, timeperiod=14)
    rsi_diff = np.nanmax(np.abs(numta_rsi - talib_rsi))
    print(f"RSI max difference: {rsi_diff}")

    # Compare MACD
    numta_macd, numta_signal, numta_hist = MACD(close)
    talib_macd, talib_signal, talib_hist = talib.MACD(close)
    macd_diff = np.nanmax(np.abs(numta_macd - talib_macd))
    print(f"MACD max difference: {macd_diff}")

Full Accuracy Suite

# Install TA-Lib (requires C library)
# See: https://github.com/TA-Lib/ta-lib-python

# Run all accuracy tests
pytest tests/accuracy/ -v --tb=short

# Run specific indicator test
pytest tests/accuracy/test_overlap_accuracy.py -v

Known Differences

Lookback Period Handling

numta uses NaN for the lookback period, matching TA-Lib behavior:

import numpy as np
from numta import SMA

close = np.array([1.0, 2.0, 3.0, 4.0, 5.0])
sma = SMA(close, timeperiod=3)
print(sma)  # [nan, nan, 2.0, 3.0, 4.0]

MAMA/FAMA (Hilbert Transform)

The MAMA indicator uses a simplified implementation that may differ slightly from TA-Lib in edge cases due to the complexity of the Hilbert Transform algorithm.

SAR Edge Cases

Parabolic SAR may show minor differences (< 1e-8) in the first few values due to initialization differences.

Reporting Accuracy Issues

If you find an accuracy discrepancy:

1. Verify you're using the same input data types (float64)
2. Check the parameters match exactly
3. Open an issue with:
4. numta version
5. TA-Lib version
6. Sample data that reproduces the issue
7. Expected vs actual output

Migration from TA-Lib

numta is designed as a drop-in replacement for TA-Lib:

# Before (TA-Lib)
import talib
sma = talib.SMA(close, timeperiod=20)
rsi = talib.RSI(close, timeperiod=14)

# After (numta)
from numta import SMA, RSI
sma = SMA(close, timeperiod=20)
rsi = RSI(close, timeperiod=14)

The function signatures and output formats are designed to be identical.