Controller Development

Phase 2: Controller Development

Translate the confirmed signal spec into a working controller — fully vectorized, uploaded, with an initial named config ready for backtesting.

Step 1 — Define the config class

Every tunable value is a config field. No magic numbers in the signal logic — a hardcoded threshold cannot be swept, which kills Phase 3.

Required fields (from the base class):

• controller_name — snake_case and descriptive (ema_cross_adx_filter)
• connector_name — e.g. binance_perpetual
• trading_pair — e.g. BTC-USDT
• candles_connector / candles_trading_pair — usually mirror the above
• interval — from research (e.g. 1h)
• max_records — max(indicator_length) * 3, minimum 200

Step 2 — Implement update_processed_data

Critical rules:

1. Fully vectorized — no iterrows(), no apply() with Python lambdas, no loops over rows. The engine reads the signal column in bulk; a row-loop silently produces wrong results.
2. The signal column holds only 1, -1 or 0 — not booleans, not floats.
3. Every tunable value comes from self.config.<param>.
4. End with exactly these two exports.

from hummingbot.strategy_v2.controllers.directional_trading_controller_base import (
    DirectionalTradingControllerBase,
    DirectionalTradingControllerConfigBase,
)
import pandas_ta as ta
import pandas as pd


class EmaCrossAdxConfig(DirectionalTradingControllerConfigBase):
    controller_name: str = "ema_cross_adx"
    connector_name: str = "binance_perpetual"
    trading_pair: str = "BTC-USDT"
    candles_connector: str = "binance_perpetual"
    candles_trading_pair: str = "BTC-USDT"
    interval: str = "1h"
    max_records: int = 500
    # ---- strategy params (one field per tunable value) ----
    ema_fast: int = 20
    ema_slow: int = 50
    adx_period: int = 14
    adx_threshold: float = 25.0


class EmaCrossAdx(DirectionalTradingControllerBase):
    async def update_processed_data(self):
        df = self.market_data_provider.get_candles_df(
            self.config.candles_connector,
            self.config.candles_trading_pair,
            self.config.interval,
            self.config.max_records,
        )

        # 1. Indicators
        df.ta.ema(length=self.config.ema_fast, append=True)
        df.ta.ema(length=self.config.ema_slow, append=True)
        df.ta.adx(length=self.config.adx_period, append=True)

        # 2. Column references
        fast_col = f"EMA_{self.config.ema_fast}"
        slow_col = f"EMA_{self.config.ema_slow}"
        adx_col  = f"ADX_{self.config.adx_period}"

        # 3. Vectorized signal logic
        long_cond  = (df[fast_col] > df[slow_col]) & (df[adx_col] > self.config.adx_threshold)
        short_cond = (df[fast_col] < df[slow_col]) & (df[adx_col] > self.config.adx_threshold)

        df["signal"] = 0
        df.loc[long_cond,  "signal"] = 1
        df.loc[short_cond, "signal"] = -1

        # 4. Export
        self.processed_data["signal"]   = int(df["signal"].iloc[-1])
        self.processed_data["features"] = df

Step 3 — Self-review checklist

Before uploading:

• No row loops — grep for iterrows, apply, for idx; must be zero
• All params from config — no hardcoded numbers in the signal logic
• Signal values — only 1, -1, 0
• Column names — pandas_ta convention: EMA_{length}, RSI_{length}, ADX_{length}, MACD_{fast}_{slow}_{signal}, BBL_{length}_{std}, BBU_{length}_{std}, SUPERTd_{length}_{multiplier}. Verify against actual output columns — a mismatch is a runtime KeyError
• NaN handling — indicators emit NaN during warm-up; NaN comparisons yield False, which correctly maps to signal = 0
• signal populated for all rows — default 0, then overwritten by masks
• iloc[-1], not iloc[0] — the current bar is what the engine acts on
• Indicator lookback < max_records - 50 — headroom for warm-up
• No look-ahead — .shift(1) wherever entry-bar logic needs it
• Complete module — both classes, all imports

Step 4 — Upload the controller

manage_controllers(
    action="upsert",
    target="controller",
    controller_type="directional_trading",
    controller_name="ema_cross_adx",
    controller_code="<full python source>",
)

Confirm created: true / updated: true. Add confirm_override=True when replacing an existing template.

The controller_code string is wrapped as a Controller object by the MCP tool — the endpoint rejects a bare source string.

A new .py may still be shadowed by the module cached in sys.modules — if the upload succeeds but describe keeps showing the old fields, restart the API container.

Step 5 — Create the initial config

manage_controllers(
    action="upsert",
    target="config",
    config_name="ema_cross_adx_v1",
    config_data={
        "controller_type": "directional_trading",
        "controller_name": "ema_cross_adx",
        "connector_name": "binance_perpetual",
        "trading_pair": "BTC-USDT",
        "candles_connector": "binance_perpetual",
        "candles_trading_pair": "BTC-USDT",
        "interval": "1h",
        "max_records": 500,
        # strategy params
        "ema_fast": 20,
        "ema_slow": 50,
        "adx_period": 14,
        "adx_threshold": 25.0,
        # execution params
        "total_amount_quote": 100,
        "stop_loss": 0.03,
        "take_profit": 0.05,
        "trailing_stop": None,
        "time_limit": 86400,
        "leverage": 5,
    },
)

Rules that actually bite:

• Numbers must be numbers, not strings. "ema_fast": "5" reaches the backtesting engine as a string and fails on arithmetic.
• controller_name must match the controller module name (ema_cross_adx → bots/controllers/directional_trading/ema_cross_adx.py). It is how the backend resolves the config class.
• config_name becomes the YAML filename and the config id.
• Config classes set extra="forbid" — a typo'd field is a hard rejection, not a warning. Discover the real field set with manage_controllers(action="describe", controller_name="<name>").
• confirm_override=True is required when overwriting an existing config.

Step 6 — Modifying an existing controller

Config only:

1. manage_controllers(action="describe", config_name="...") — read current values
2. Re-upsert with confirm_override=True and the updated config_data

Controller code:

1. manage_controllers(action="describe", controller_name="...", include_code=True)
2. Rewrite the full update_processed_data — never patch partial snippets
3. Re-run the self-review checklist
4. Upsert with confirm_override=True
5. Re-upload any configs referencing it — they do not auto-update

Go/No-Go → Phase 3

✅ GO — controller and config uploaded without errors, checklist fully passes, code is a complete importable module.

❌ NO-GO — any row-level loop, hardcoded magic numbers in the signal logic, column name mismatch, or max_records < 2× the longest indicator length.

Artifacts

1. Controller source (complete module)
2. Named config with all parameters
3. Self-review checklist results