How to determine the best deviation_model for a PyTechnicalIndicators function
This guide shows how to programmatically determine the best deviation_model for your indicator using the Python package PyTechnicalIndicators.
The rating model is overly simplified and should be refined to suit your needs before usage.
🎯 Goal
- Determine the best
deviation_modelfor the Moving Constant Bands from a year of data
This guide assumes you can load price data from a CSV file (see the "Get data from CSV" section below).
📦 Requirements
Install PyTechnicalIndicators:
pip install pytechnicalindicators
💻 Step-by-Step
1. Get data from CSV
This example expects a CSV file with either
- a header containing a "close" column (case-insensitive), or
- a single column of prices (no header)
Example loader:
import csv
def load_prices_from_csv(path: str) -> list[float]:
prices: list[float] = []
with open(path, "r", newline="") as f:
# Try to read as DictReader first (with a header)
sample = f.read(4096)
f.seek(0)
has_header = "," in sample and any(ch.isalpha() for ch in sample.splitlines()[0])
if has_header:
reader = csv.DictReader(f)
# Find a 'close' column in a case-insensitive way
close_key = None
if reader.fieldnames:
for k in reader.fieldnames:
if k and k.lower() == "close":
close_key = k
break
if close_key is None and reader.fieldnames:
# Fallback to the first column if 'close' not found
close_key = reader.fieldnames[0]
for row in reader:
try:
prices.append(float(row[close_key]))
except (ValueError, TypeError, KeyError):
continue
else:
# No header: read first column only
f.seek(0)
reader = csv.reader(f)
for row in reader:
if not row:
continue
try:
prices.append(float(row[0]))
except ValueError:
continue
return prices
2. Calculate moving constant bands for multiple deviation models
The default deviation model is the standard deviation; however other models may provide more insight.
We’ll test several deviation models while keeping the constant model fixed (e.g., exponential moving average), a deviation multiplier (e.g., 2.0), and a short period (e.g., 5) for illustration.
from pytechnicalindicators import candle_indicators as ci
period = 5
deviation_multiplier = 2.0
constant_model = "exponential" # or "exponential_moving_average"
deviation_models = [
"standard", # or "standard_deviation"
"mean", # or "mean_absolute_deviation"
"median", # or "median_absolute_deviation"
"mode", # or "mode_absolute_deviation"
"ulcer", # or "ulcer_index"
]
# Example:
# bands = ci.bulk.moving_constant_bands(prices, constant_model, dev_model, deviation_multiplier, period)
3. Rate each model to find the best
The logic is overly simple for the purpose of the guide.
- If price > upper band (overbought) and next price < current price, model gets +1
- If price < lower band (oversold) and next price > current price, model gets +1
We normalize the score by the number of "attempts" (how many times we evaluated a signal).
Important: moving_constant_bands returns tuples ordered as (lower_band, moving_constant, upper_band). Align the band window to the price index i using bands_idx = i - (period - 1).
def choose_best_deviation_model(
prices: list[float],
constant_model_type: str = "exponential",
deviation_multiplier: float = 2.0,
period: int = 5,
) -> tuple[str, float]:
from pytechnicalindicators import candle_indicators as ci
deviation_models = [
"standard", # or "standard_deviation"
"mean", # or "mean_absolute_deviation"
"median", # or "median_absolute_deviation"
"mode", # or "mode_absolute_deviation"
"ulcer", # or "ulcer_index"
]
best_rating = -1.0
best_model = deviation_models[0]
for dm in deviation_models:
bands = ci.bulk.moving_constant_bands(
prices,
constant_model_type,
dm,
deviation_multiplier,
period,
)
# bands[k] corresponds to price index i = k + (period - 1)
current_rating = 0.0
attempts = 0.0
for i in range(period - 1, len(prices) - 1):
bands_idx = i - (period - 1)
lower_band, _, upper_band = bands[bands_idx]
# Overbought: expect next price to fall
if prices[i] > upper_band:
attempts += 1.0
if prices[i + 1] < prices[i]:
current_rating += 1.0
# Oversold: expect next price to rise
if prices[i] < lower_band:
attempts += 1.0
if prices[i + 1] > prices[i]:
current_rating += 1.0
average_rating = (current_rating / attempts) if attempts > 0 else 0.0
if average_rating > best_rating:
best_rating = average_rating
best_model = dm
return best_model, best_rating
4. Full example
import sys
# Reuse load_prices_from_csv and choose_best_deviation_model from above
def main():
if len(sys.argv) < 2:
print("Usage: python choose_deviation_model.py <path_to_csv>")
sys.exit(1)
csv_path = sys.argv[1]
prices = load_prices_from_csv(csv_path)
print(f"Loaded {len(prices)} prices")
if len(prices) < 5:
print("Not enough data to compute moving constant bands with period 5.")
sys.exit(1)
best_model, best_rating = choose_best_deviation_model(
prices,
constant_model_type="exponential", # keep constant model fixed for this search
deviation_multiplier=2.0,
period=5,
)
print(f"Best model for moving constant bands is {best_model} with a rating of {best_rating}")
if __name__ == "__main__":
main()
🧪 Output
A runnable sample of the code can be found in choose_deviation_model.py
Loaded 251 prices
Best model for moving constant bands is median with a rating of 0.6666666666666666
✅ Next Steps
- Programmatically choose a period
- Programmatically choose a
constant_model_type - Programmatically choose a deviation multiplier
- Combine all selections
- Introduce the notion of punishment to the rating system (e.g., penalize false signals/whipsaws)