💰 The plan to optimize your strategies

[tripolskypetr]

Volatility-Filtered EMA Strategy

GARCH is the ATR without magic constant multiply coef. It will read candle history and predict t+1 candle volatility

#77 (comment)

Responsibility:

• Signal Generation: pinets
• Signal Filtering: backtest-kit, garch, json-inference

The Problem

EMA crossovers generate too many signals. Most of them occur during low-volatility sideways markets where price doesn't move enough to cover fees and slippage. The result: death by a thousand cuts.

The fix: before entering any trade, ask the model — "is there enough expected movement to justify this entry?"

Architecture

flowchart TD
    A[4h Candles] -->|EMA trend direction| B{LONG or SHORT?}
    B -->|direction decided| C[15m Candles]
    C -->|Pine Script| D[Entry Signal]
    D --> E{predict filter}
    E -->|sigma too low or unreliable| F[Skip]
    E -->|sigma covers costs| G[Enter Position]
    G --> H[Partial Profit via Kelly]
    H --> I[Trailing Stop after Breakeven]
    I --> J{4h timeout reached?}
    J -->|no| I
    J -->|yes| K[Close Remaining Position]

Loading

Step 1 — Trend Direction (4h)

The higher timeframe decides one thing: do we only look for longs, or only for shorts.

• Input: 200+ candles on the 4h chart
• Method: EMA crossover (or any trend indicator — MACD, Supertrend, etc.)
• Output: a single bias — LONG or SHORT

No entries happen on this timeframe. It only sets the filter for step 2.

Why 4h: short enough to catch regime changes within days, long enough to avoid whipsaws. Daily works too but reacts slower.

Step 2 — Entry Signal (15m)

Pine Script runs on the 15m chart and generates candidate entry points in the direction decided by step 1.

• If 4h says LONG — only bullish signals pass
• If 4h says SHORT — only bearish signals pass

This already eliminates ~50% of false signals (counter-trend entries).

Why 15m: gives precise entries within the 4h trend. Lower timeframes (1m, 5m) have too much microstructure noise. Higher (1h) lose precision.

Step 3 — Volatility Filter (GARCH)

This is where the predict function comes in. Before executing any entry signal from step 2, we check if the expected price movement justifies the trade.

flowchart LR
    A[Entry Signal] --> B["predict(candles, '15m')"]
    B --> C{reliable?}
    C -->|false| D[Skip — model inadequate]
    C -->|true| E{sigma * price > fees + slippage + min_profit?}
    E -->|no| F[Skip — not enough movement]
    E -->|yes| G[Execute Entry]

Loading

What predict does

1. Takes 300+ candles on 15m
2. Auto-selects GARCH or EGARCH based on leverage effect
3. Returns sigma — expected price movement for the next candle (as a fraction)
4. Returns reliable — whether the model converged and passed adequacy tests

The filter logic

expected_move = sigma * current_price
total_cost = slippage + exchange_fee_open + exchange_fee_close

if (!reliable) → skip
if (expected_move < total_cost + minimum_profit) → skip
else → enter

minimum_profit is your decision, not the model's. It represents the smallest gain worth your time and risk. Without it, you enter trades with zero expected value after costs.

What this catches

Scenario	sigma	Action	Why
Tight consolidation	0.1–0.3%	Skip	Price won't move enough to cover fees
Normal trend	0.5–1.5%	Enter	Enough room for TP with acceptable SL
High volatility	>2%	Enter with caution	Wide SL required, but TP potential is large
Model failure	any	Skip	reliable = false — data doesn't fit GARCH

Step 4 — Position Management

Once inside a trade, three rules govern the exit:

Partial profit (Kelly)

Take partial profit at a Kelly-optimal fraction of the position. Kelly criterion:

f* = (win_rate * avg_win - (1 - win_rate) * avg_loss) / avg_win

Use half-Kelly (f*/2) until you have 100+ trades for stable win rate estimation. Full Kelly is mathematically optimal but practically fragile — overestimates edge when sample size is small.

Recalculate Kelly parameters every N trades as your sample grows.

Trailing stop after breakeven

After partial profit is taken:

1. Move stop-loss to breakeven (entry price + fees)
2. Trail the stop behind price as it moves in your favor
3. Let the remaining position ride the trend

This converts a winning trade into a risk-free position. The worst outcome after partial profit is breaking even on the remainder.

Timeout: 16 candles

If TP is not reached within 16 candles of 15m (= 4 hours), close the remaining position at market.

Why 16 candles, not 4 hours: aligns with your entry timeframe. If you entered on candle N, you exit on candle N+16 regardless of wall clock time. Avoids edge cases where entering at the end of a 4h window gives you only minutes.

Why 4 hours at all: matches the higher timeframe. If the 4h trend is valid, a 15m entry should reach TP within one 4h candle. If it doesn't — the signal was likely noise.

flowchart TD
    A[Position Opened] --> B[Price hits partial TP?]
    B -->|yes| C[Take Partial Profit]
    C --> D[Move SL to Breakeven]
    D --> E[Trail Stop]
    E --> F{TP reached?}
    F -->|yes| G[Close Full Position]
    F -->|no| H{16 candles passed?}
    H -->|no| E
    H -->|yes| I[Close Remaining at Market]
    B -->|no| H

Loading

What This Strategy Does NOT Do

• Does not predict direction. GARCH says "price can move ±1.2%", not "price will go up." Direction comes from EMA.
• Does not replace risk management. Position sizing, max drawdown, correlation between positions — all separate concerns.
• Does not guarantee profits. It filters out entries where the math doesn't work. The remaining entries still depend on your directional accuracy.

Summary

Step	Timeframe	Tool	Purpose
1	4h	EMA	Trend bias (long/short)
2	15m	Pine Script	Entry signal
3	15m	predict()	Volatility filter — is there enough movement?
4	15m	Kelly + trailing	Position management — partial profit, trail, timeout

The core insight: EMA tells you where, GARCH tells you whether it's worth going there.

[tripolskypetr]

So here is the plan to make it all work together

#77 (comment)

@alaa-eddine

[tripolskypetr]

GARCH is impossible in Pine Script because it requires iterative maximum likelihood optimization with unknown convergence steps, matrix operations for gradient descent, and dynamic parameter estimation from data—none of which Pine Script supports.

//@version=5
indicator("ATR with magic constant", overlay=false)

// Magic constant — why 14? No statistical justification
ATR_LENGTH = 14  // magic

atrValue = ta.atr(ATR_LENGTH)
plot(atrValue, "ATR", color.orange)

Unlike EWMA or ATR which require manually chosen "magic constants" (smoothing factor, period length), GARCH estimates its parameters directly from the data through optimization, making it statistically rigorous but computationally demanding beyond Pine Script's capabilities.

//@version=5
indicator("EWMA Volatility", overlay=false)

// Magic constant — lambda 0.94 from RiskMetrics 1996
// Why not 0.93 or 0.95? Because JP Morgan said so
LAMBDA = 0.94  // magic

returns = math.log(close / close[1])
var float ewmaVar = 0.0
ewmaVar := LAMBDA * nz(ewmaVar[1]) + (1 - LAMBDA) * math.pow(returns, 2)

plot(math.sqrt(ewmaVar) * 100, "EWMA Vol %", color.purple)

P.S. Monte Carlo optimization randomly samples parameter combinations from the search space, which is slow. GARCH takes 500ms to self-prepare in runtime

[alaa-eddine]

hey @tripolskypetr Thank you very much for this awesome contribution.
You have interesting concepts there that many algo-traders would certainly find valuable.

Please note that I edited the discussion title to not look like a financial advice, since the main goal of this repo is to provide technical solutions for educational purposes.

[tripolskypetr]

Ok, got it

Could you look at #112, #111

@alaa-eddine

[pegaltier]

Hey @tripolskypetr I can't agree more with @alaa-eddine! You have impressive skills, the javascript community is lucky to have you, sharing your knowledge. I will try to connect your garch lib to my system and I have added your latest libraries to my awesome dev list, see commit here: pegaltier/awesome-utils-dev@aed7eb0