Technical Indicators, Tested Honestly¶

Run the 50-indicator engine on real data and measure which signals actually correlate with forward returns.

Part 27 of 36 in the ServLoci algo/options trading notebook series — full index in notebooks/README.md.

Setup — no broker account needed¶

!pip install -q yfinance

import yfinance as yf
import pandas as pd
import numpy as np

# No broker account, no API key, no ServLoci setup needed for this chapter —
# yfinance reads public end-of-day data from Yahoo Finance. Indian tickers take
# an ".NS" suffix (NSE) or ".BO" (BSE); index tickers are prefixed with "^"
# (^NSEI = Nifty 50, ^BSESN = Sensex, ^GSPC = S&P 500).
NSE_TICKER = "RELIANCE.NS"
US_TICKER = "AAPL"
INDEX_TICKER = "^NSEI"

print("yfinance", yf.__version__)

"This indicator works" is a claim, not a fact¶

Search for any popular indicator and you'll find one video insisting it's essential and the next calling it useless — both presented with total confidence, neither showing a single line of code. Notebook 21 built all 50 indicators and described what each family measures; it stopped short of asking the harder question: on real data, did any of them actually correlate with what happened next?

This chapter asks that question properly — not to hand you a verdict to memorize, but to show you how to interrogate a claim yourself instead of trusting whoever sounds most confident. The code below computes indicators on real pulled history and tests a few textbook claims against real forward returns.

!pip install -q pandas numpy matplotlib
import numpy as np
import pandas as pd

def compute_top_50(frame):
    """Return 50 named indicators from an OHLCV DataFrame.

    Input columns are case-insensitive: open, high, low, close and volume.
    Warm-up rows contain NaN by design; never backfill them into a live signal.
    """
    df = frame.rename(columns={c: str(c).lower() for c in frame.columns}).copy()
    required = {"open", "high", "low", "close", "volume"}
    missing = required.difference(df.columns)
    if missing:
        raise ValueError(f"missing OHLCV columns: {sorted(missing)}")
    o, h, l, c, v = (df[x].astype(float) for x in ("open", "high", "low", "close", "volume"))
    out = pd.DataFrame(index=df.index)
    safe = lambda x: x.replace([np.inf, -np.inf], np.nan)
    ema = lambda x, n: x.ewm(span=n, adjust=False, min_periods=n).mean()
    wma = lambda x, n: x.rolling(n).apply(
        lambda a: np.dot(a, np.arange(1, n + 1)) / (n * (n + 1) / 2), raw=True
    )

    # Trend and moving-average family (1-11)
    out["01_sma_20"] = c.rolling(20).mean()
    out["02_ema_20"] = ema(c, 20)
    out["03_wma_20"] = wma(c, 20)
    out["04_hma_20"] = wma(2 * wma(c, 10) - wma(c, 20), 4)
    e1 = ema(c, 20); e2 = ema(e1, 20); e3 = ema(e2, 20)
    out["05_dema_20"] = 2 * e1 - e2
    out["06_tema_20"] = 3 * e1 - 3 * e2 + e3
    out["07_vwma_20"] = safe((c * v).rolling(20).sum() / v.rolling(20).sum())
    macd = ema(c, 12) - ema(c, 26)
    out["08_macd"] = macd
    out["09_macd_signal"] = ema(macd, 9)
    out["10_ppo"] = safe(100 * macd / ema(c, 26))
    ex1 = ema(c, 15); ex2 = ema(ex1, 15); ex3 = ema(ex2, 15)
    out["11_trix"] = ex3.pct_change(fill_method=None) * 100

    # Momentum and oscillator family (12-24)
    out["12_roc_12"] = c.pct_change(12, fill_method=None) * 100
    out["13_momentum_10"] = c.diff(10)
    delta = c.diff(); gain = delta.clip(lower=0); loss = -delta.clip(upper=0)
    avg_gain = gain.ewm(alpha=1/14, adjust=False, min_periods=14).mean()
    avg_loss = loss.ewm(alpha=1/14, adjust=False, min_periods=14).mean()
    out["14_rsi_14"] = 100 - (100 / (1 + safe(avg_gain / avg_loss)))
    low14, high14 = l.rolling(14).min(), h.rolling(14).max()
    stoch = safe(100 * (c - low14) / (high14 - low14))
    out["15_stochastic_k"] = stoch
    out["16_stochastic_d"] = stoch.rolling(3).mean()
    out["17_williams_r"] = safe(-100 * (high14 - c) / (high14 - low14))
    typical = (h + l + c) / 3
    tp_mean = typical.rolling(20).mean()
    mean_dev = typical.rolling(20).apply(lambda a: np.mean(np.abs(a - a.mean())), raw=True)
    out["18_cci_20"] = safe((typical - tp_mean) / (0.015 * mean_dev))
    prev = c.shift(1)
    buy_pressure = c - pd.concat([l, prev], axis=1).min(axis=1)
    true_range = pd.concat([h - l, (h - prev).abs(), (l - prev).abs()], axis=1).max(axis=1)
    out["19_ultimate_oscillator"] = safe(100 * (
        4 * buy_pressure.rolling(7).sum() / true_range.rolling(7).sum()
        + 2 * buy_pressure.rolling(14).sum() / true_range.rolling(14).sum()
        + buy_pressure.rolling(28).sum() / true_range.rolling(28).sum()
    ) / 7)
    midpoint = (h + l) / 2
    out["20_awesome_oscillator"] = midpoint.rolling(5).mean() - midpoint.rolling(34).mean()
    r1, r2, r3, r4 = (c.pct_change(n, fill_method=None) * 100 for n in (10, 15, 20, 30))
    out["21_kst"] = r1.rolling(10).sum() + 2*r2.rolling(10).sum() + 3*r3.rolling(10).sum() + 4*r4.rolling(15).sum()
    pc = c.diff(); apc = pc.abs()
    out["22_tsi"] = safe(100 * ema(ema(pc, 25), 13) / ema(ema(apc, 25), 13))
    streak = pd.Series(0.0, index=c.index)
    for i in range(1, len(c)):
        direction = np.sign(c.iloc[i] - c.iloc[i - 1])
        prior = streak.iloc[i - 1]
        streak.iloc[i] = 0 if direction == 0 else direction * (abs(prior) + 1 if np.sign(prior) == direction else 1)
    def rsi_series(x, n):
        d = x.diff(); g = d.clip(lower=0).ewm(alpha=1/n, adjust=False, min_periods=n).mean()
        q = (-d.clip(upper=0)).ewm(alpha=1/n, adjust=False, min_periods=n).mean()
        return 100 - 100 / (1 + safe(g / q))
    pct_rank = c.pct_change(fill_method=None).rolling(100).apply(lambda a: 100 * (a[-1] > a[:-1]).mean(), raw=True)
    out["23_connors_rsi"] = (rsi_series(c, 3) + rsi_series(streak, 2) + pct_rank) / 3
    sum_gain, sum_loss = gain.rolling(14).sum(), loss.rolling(14).sum()
    out["24_cmo_14"] = safe(100 * (sum_gain - sum_loss) / (sum_gain + sum_loss))

    # Volatility and channel family (25-36)
    mid = c.rolling(20).mean(); std = c.rolling(20).std(ddof=0)
    upper, lower = mid + 2*std, mid - 2*std
    out["25_bollinger_upper"] = upper
    out["26_bollinger_lower"] = lower
    out["27_bollinger_bandwidth"] = safe(100 * (upper - lower) / mid)
    atr = true_range.ewm(alpha=1/14, adjust=False, min_periods=14).mean()
    out["28_atr_14"] = atr
    out["29_natr_14"] = safe(100 * atr / c)
    out["30_true_range"] = true_range
    kel_mid = ema(c, 20)
    out["31_keltner_upper"] = kel_mid + 2 * atr
    out["32_keltner_lower"] = kel_mid - 2 * atr
    out["33_donchian_upper"] = h.rolling(20).max()
    out["34_donchian_lower"] = l.rolling(20).min()
    out["35_stddev_20"] = std
    out["36_historical_volatility"] = np.log(c / c.shift(1)).rolling(20).std(ddof=0) * np.sqrt(252) * 100

    # Directional, stop and cloud family (37-46)
    up_move, down_move = h.diff(), -l.diff()
    plus_dm = up_move.where((up_move > down_move) & (up_move > 0), 0.0)
    minus_dm = down_move.where((down_move > up_move) & (down_move > 0), 0.0)
    plus_di = safe(100 * plus_dm.ewm(alpha=1/14, adjust=False, min_periods=14).mean() / atr)
    minus_di = safe(100 * minus_dm.ewm(alpha=1/14, adjust=False, min_periods=14).mean() / atr)
    out["37_adx_14"] = safe(100 * (plus_di - minus_di).abs() / (plus_di + minus_di)).ewm(alpha=1/14, adjust=False, min_periods=14).mean()
    out["38_plus_di"] = plus_di
    out["39_minus_di"] = minus_di
    out["40_aroon_up"] = h.rolling(25).apply(lambda a: 100 * (np.argmax(a) + 1) / len(a), raw=True)
    out["41_aroon_down"] = l.rolling(25).apply(lambda a: 100 * (np.argmin(a) + 1) / len(a), raw=True)
    out["42_vortex_plus"] = safe((h - l.shift(1)).abs().rolling(14).sum() / true_range.rolling(14).sum())
    out["43_vortex_minus"] = safe((l - h.shift(1)).abs().rolling(14).sum() / true_range.rolling(14).sum())
    psar = pd.Series(np.nan, index=c.index)
    if len(c):
        bull, af, extreme = True, 0.02, h.iloc[0]
        psar.iloc[0] = l.iloc[0]
        for i in range(1, len(c)):
            candidate = psar.iloc[i-1] + af * (extreme - psar.iloc[i-1])
            if bull:
                candidate = min(candidate, l.iloc[i-1], l.iloc[max(i-2, 0)])
                if l.iloc[i] < candidate: bull, candidate, af, extreme = False, extreme, 0.02, l.iloc[i]
                elif h.iloc[i] > extreme: extreme, af = h.iloc[i], min(af + 0.02, 0.2)
            else:
                candidate = max(candidate, h.iloc[i-1], h.iloc[max(i-2, 0)])
                if h.iloc[i] > candidate: bull, candidate, af, extreme = True, extreme, 0.02, h.iloc[i]
                elif l.iloc[i] < extreme: extreme, af = l.iloc[i], min(af + 0.02, 0.2)
            psar.iloc[i] = candidate
    out["44_parabolic_sar"] = psar
    out["45_ichimoku_conversion"] = (h.rolling(9).max() + l.rolling(9).min()) / 2
    out["46_ichimoku_base"] = (h.rolling(26).max() + l.rolling(26).min()) / 2

    # Volume and money-flow family (47-50)
    out["47_obv"] = (np.sign(c.diff()).fillna(0) * v).cumsum()
    raw_money = typical * v; positive = raw_money.where(typical.diff() > 0, 0); negative = raw_money.where(typical.diff() < 0, 0)
    out["48_mfi_14"] = 100 - 100 / (1 + safe(positive.rolling(14).sum() / negative.rolling(14).sum()))
    money_flow_multiplier = safe(((c - l) - (h - c)) / (h - l))
    money_flow_volume = money_flow_multiplier * v
    out["49_cmf_20"] = safe(money_flow_volume.rolling(20).sum() / v.rolling(20).sum())
    out["50_accumulation_distribution"] = money_flow_volume.fillna(0).cumsum()
    assert out.shape[1] == 50
    return out

hist = yf.Ticker(INDEX_TICKER).history(period="5y", interval="1d")
hist = hist.dropna(subset=["Open", "High", "Low", "Close", "Volume"])
print(f"{INDEX_TICKER}: {len(hist)} daily bars, {hist.index.min().date()} to {hist.index.max().date()}")

indicators = compute_top_50(hist)  # column names are case-insensitive on the way in
close = hist["Close"]
print("indicator columns available:", indicators.shape[1])

Testing three textbook claims¶

RSI(14): "below 30 is oversold, expect a bounce; above 70 is overbought, expect a pullback."
MACD: "a bullish crossover (MACD line crossing above its signal line) marks a shift toward upward momentum."
ADX(14): "above ~25 the market is trending, and trend-following setups do better; below ~20 it's chop."

Each is testable: bucket historical days by the condition, then look at what actually happened over the following N trading days. Note the direction of time here — this is a retrospective research question ("did this condition tend to precede that outcome"), which is legitimate to compute with .shift(-horizon) on data you already have in front of you. That is different from lookahead bias inside a backtest, where a strategy's simulated decision on day T secretly uses information only available after day T. Notebook 29 draws that line precisely — keep it in mind here, because it's easy to blur the two.

horizon = 10  # trading days
forward_return = close.shift(-horizon) / close - 1

rsi = indicators["14_rsi_14"]
oversold = forward_return[rsi < 30]
overbought = forward_return[rsi > 70]
print(f"RSI<30  — {oversold.count():4d} obs, mean {horizon}d forward return {oversold.mean():+.2%}, std {oversold.std():.2%}")
print(f"RSI>70  — {overbought.count():4d} obs, mean {horizon}d forward return {overbought.mean():+.2%}, std {overbought.std():.2%}")

macd, signal = indicators["08_macd"], indicators["09_macd_signal"]
bullish_cross = (macd.shift(1) < signal.shift(1)) & (macd > signal)
cross_fwd = forward_return[bullish_cross]
print(f"MACD bullish crossover — {cross_fwd.count():4d} events, mean {horizon}d forward return {cross_fwd.mean():+.2%}, std {cross_fwd.std():.2%}")

adx = indicators["37_adx_14"]
trending = forward_return[adx > 25]
choppy = forward_return[adx < 20]
print(f"ADX>25 (trending) — {trending.count():4d} obs, mean {horizon}d forward return {trending.mean():+.2%}, std {trending.std():.2%}")
print(f"ADX<20 (choppy)   — {choppy.count():4d} obs, mean {horizon}d forward return {choppy.mean():+.2%}, std {choppy.std():.2%}")

What this test does — and doesn't — prove¶

Whatever numbers come out above, treat them as a starting point, not a verdict:

One ticker, one history. These 5 years of Nifty behavior are one sample path out of many that could have happened. A pattern here may be specific to this index, this regime (rate cycle, liquidity conditions), or pure coincidence.
Overlapping windows aren't independent observations. A 10-day forward return computed on every single day overlaps with the 9 days before and after it — the "count" in each bucket wildly overstates how much independent evidence you actually have. Proper statistical testing would need to account for this (block bootstrap, or non-overlapping samples), which this chapter deliberately doesn't do, so as not to imply a rigor it doesn't have.
No significance test, no multiple-comparison correction. We tested three claims; if you tested thirty, some would look good by chance alone even with zero real edge. That's a preview of the overfitting trap notebook 29 demonstrates deliberately.
A positive average is necessary, not sufficient. Even a real, non-random edge has to survive transaction costs, slippage, and regime change before it's tradeable — none of which this notebook has applied yet.

The honest conclusion from a chapter like this is almost never "indicator X works" or "indicator X is useless." It's "here's what the data shows, here's how much I should trust it, and here's what I'd need to check next" — which is a far more useful habit than picking a side in a comments-section argument.

« Previous: Fundamental Analysis with Real Filings
Next: Options, Priced Against Reality »

Try the concepts above interactively: Options Strategy Builder · Docs · Get your static IP