Indicator Best Practices: Standardization

Raw values are hard to threshold. Standardized values tell you how unusual the current reading is — which is what actually matters.

The problem with raw indicator values

Suppose you want to build a strategy that goes long when "buying pressure is high." You use net taker volume as your signal. The raw value might be 8,420 BTC in a quiet week and 85,000 BTC during a volatile period. What threshold do you set? Any fixed number will be too sensitive half the time and too slow the other half.

The same problem shows up with price-derived indicators: an SMA difference of 500 USDT means something very different at $20,000 BTC vs $90,000 BTC.

Standardization converts a raw value into a question: "How far is this from normal, measured in standard deviations?" This makes the reading comparable across different market conditions and different symbols.

Rolling Z-score: the most useful method

A Z-score measures how many standard deviations the current value is from the recent mean:

z = (value - rolling_mean) / rolling_std

With a 30-day rolling window, a Z-score of +2.0 means the current reading is 2 standard deviations above the past month's average — unusual, but not extreme. A Z-score of +3.5 is extremely high by recent standards.

Z-scores have a useful property: they're roughly unit-free and comparable across symbols. A TI Z-score of +2.0 on BTCUSDT means roughly the same thing as a TI Z-score of +2.0 on ETHUSDT — both are unusually high relative to their own recent history.

In a strategy, implementing a rolling Z-score looks like this:

window = 720  # 30 days × 24h for 1h interval
df['z'] = (
    (df['raw_indicator'] - df['raw_indicator'].rolling(window).mean())
    / df['raw_indicator'].rolling(window).std()
)

Blave alpha indicators are already standardized

You don't need to standardize Blave's proprietary indicators yourself. They are already delivered as normalized scores comparable across symbols and time periods:

This is why threshold strategies using Blave indicators use values like ENTRY_TH = 1.693 rather than a raw volume number — that threshold has a consistent meaning: "enter when TI is 1.7 standard deviations above its recent average."

When you do need to standardize yourself

Not all indicators need standardization. An SMA crossover signal (SMA_fast > SMA_slow) is already binary — there's nothing to standardize. The issue arises when you use an indicator's magnitude as a threshold.

Example 1 — SMA difference as a threshold

Using the raw price-minus-SMA as a signal breaks across price regimes:

# ✗ Breaks when BTC moves from $20k to $90k
signal[df['Close'] - df['SMA_50'] > 500] = 1.0

# ✓ Z-score of the same difference stays meaningful
diff = df['Close'] - df['SMA_50']
df['diff_z'] = (diff - diff.rolling(720).mean()) / diff.rolling(720).std()
signal[df['diff_z'] > 1.5] = 1.0

Example 2 — MACD across multiple symbols

MACD values are in price units — they're not comparable across different symbols:

# ✗ MACD = 800 means very different things on BTCUSDT vs 2330.TW
signal[df['MACD'] > 800] = 1.0

# ✓ Standardize per symbol before using as threshold
df['macd_z'] = (df['MACD'] - df['MACD'].rolling(720).mean()) / df['MACD'].rolling(720).std()
signal[df['macd_z'] > 1.0] = 1.0

Example 3 — Combining two indicators

RSI (0–100) and raw OI (billions of USD) can't be combined directly. Standardize both first:

# ✗ OI drowns out RSI entirely
combined = df['RSI'] + df['OI']

# ✓ Both as Z-scores, then combine
rsi_z = (df['RSI'] - df['RSI'].rolling(720).mean()) / df['RSI'].rolling(720).std()
oi_z  = (df['OI']  - df['OI'].rolling(720).mean())  / df['OI'].rolling(720).std()
combined = rsi_z + oi_z
signal[combined > 1.5] = 1.0

Example 4 — Shareholder count (Taiwan stocks)

The number of shareholders in a given bracket (e.g., holders of 1,000–5,000 shares) differs massively between large-cap and small-cap stocks. TSMC may have hundreds of thousands of retail holders; a small-cap stock might have a few thousand. The raw count cannot be used as a threshold across different stocks.

What matters is the change relative to that stock's own recent history — is retail participation increasing unusually fast? Standardize the count (or its rate of change) against each stock's rolling window:

# Raw retail holder count is not comparable across stocks
# ✗ meaningless as a cross-stock threshold
signal[df['retail_holders'] > 50000] = 1.0

# ✓ Standardize per stock: how unusual is the current count?
df['holders_z'] = (
    (df['retail_holders'] - df['retail_holders'].rolling(52).mean())
    / df['retail_holders'].rolling(52).std()   # 52 weeks rolling window
)
# Z-score > 2: retail unusually crowded → potential contrarian short signal
signal[df['holders_z'] > 2.0] = 0.0
signal[df['holders_z'] < -1.0] = 1.0

The same principle applies to margin balance (margin_balance), foreign institutional net buy/sell (foreign_net), or any other Taiwan stock data field that varies in absolute scale by company size.

Other standardization methods

For most Type A strategies, rolling Z-score is the right default. Percentile rank is useful when you want a bounded signal (e.g., "top 10% of historical readings") without being affected by extreme outliers that can skew a Z-score.