Ouroboros — multi-timeframe cTrader cBot

A portfolio piece by Nuraveda Lab. Mirrored on GitHub · GitLab · Codeberg.

Six ML bots — viper (M5), hydra (M1), mamba (M15), taipan (M30), cobra (H1), anaconda (H4) — merged into one cBot through a data-driven whitelist derived from 3,253 closed demo trades. Open-sourced as a portfolio piece showing how to take a multi-bot ML signal layer and ship it as a single customer-side .algo for the cTrader platform.

Free to use. Trade on a demo first; this is educational software, not investment advice. See LICENSE for the full disclaimer.

What it does

Each timeframe has its own strategy (RSI momentum, Bollinger-band reversion, session-quality bias, MA-cross trend follower). Strategies run independently on the symbols they have a proven historical edge on. When two or more bots fire the same direction on the same symbol inside a 30-min window, the position sizer scales the trade up — agreement multiplies size, not entry permission, which is what the data actually showed.

Customer-side execution — the trader runs this in their own cTrader Desktop on their broker account. Conservative drawdown caps are configurable.

What's inside

Bot	Timeframe	Strategy	Tradeable symbols (data-driven whitelist)
hydra	M1	mamba_reversion	JPN225, GER40, UK100, ETHUSD, XAUUSD, XAGUSD, EURJPY, EURUSD, AUDUSD, NZDUSD
viper	M5	momentum_hunter	BTCUSD (only — bled $838k everywhere else)
mamba	M15	mamba_reversion	JPN225 (probation: small sample, 64% WR)
taipan	M30	session_analyst	GER40, XAUUSD, UK100, JPN225 (trend tier)
cobra	H1	trend_follower	GER40 (probation: small sample, 71% WR)
anaconda	H4	—	disabled (19% WR, structurally broken)

Whitelist generated from the live PostgreSQL ml_trades table — 3,253 closed demo trades (2026-04-13 → 2026-05-12). Cells with ≥30 trades, +PnL and ≥45% WR are Full tier. High-sample low-WR positive-tail strategies (viper-BTCUSD, taipan-JPN225) are Trend tier (wider take-profit). Encouraging-but-thin cells (mamba-JPN225, cobra-GER40) are Probation (½ size).

Signal aggregation rule

The data shows multi-bot agreement does not improve win-rate, but does multiply average PnL 4–9×. Therefore:

• Any whitelisted bot may enter on its own (no agreement gate).
• Position size is multiplied by the number of OTHER bots that fired the same direction on the same symbol within the last 30 min:
  • 0 agree → 1.0× base
  • 1 agree → 1.5× base
  • 2 agree → 2.0× base
  • 3+ agree → 2.5× base (cap)

Probation tier halves all of the above. There is no hierarchical TF veto — H4 was the worst-performing TF in the data and would block profitable lower-TF trades.

Project layout

cbot/
├── Ouroboros.sln
└── Ouroboros/
    ├── Ouroboros.csproj
    ├── Ouroboros.cs               -- Robot class, parameters, OnStart, routing
    ├── Core/
    │   ├── Signal.cs              -- shared DTO + Vote enum
    │   ├── Whitelist.cs           -- compile-time bot×symbol×tier table
    │   ├── SignalBus.cs           -- rolling 30-min agreement memory
    │   ├── PositionSizer.cs       -- balance×risk%×tier×agreement → units
    │   └── RiskGuard.cs           -- daily/total DD caps + flatten switch
    └── Strategies/
        ├── IStrategy.cs           -- interface (Attach, OnBar→Signal)
        ├── MomentumHunter.cs      -- viper (RSI 48/52 cross + EMA + volume)
        ├── MambaReversion.cs      -- hydra & mamba (BB fade in ADX<25 ranges)
        ├── SessionAnalyst.cs      -- taipan (London/NY/overlap session + EMA slope)
        └── TrendFollower.cs       -- cobra (SMA9/EMA21 cross + ADX>15)

Build

Local sanity-check with the .NET 8 SDK:

# from the repo root
dotnet build Ouroboros.sln -c Release

Output: Ouroboros/bin/Release/net6.0/cbot.algo — that's the file you import into cTrader Desktop (Automate → cBots → Add → Import).

In normal use, customers don't run dotnet build. They drop the .algo (or the source folder) into their cTrader Desktop's cAlgo/Sources/Robots/ directory and let cTrader compile it.

Default parameters (cTrader UI groups)

General — Label (positions tagged <Label>-<bot>-<symbol>)

Risk (conservative drawdown presets)

• Base risk per trade % = 0.5
• Max daily drawdown % = 3.5 (leaves headroom under a 4% account rule)
• Max total drawdown % = 5.5 (leaves headroom under a 6% account rule)
• Flatten at daily DD % = 4.0 (panic-flatten before the daily cap is breached)
• Max concurrent positions = 4

Execution

• Min confidence to enter = 0.65
• ATR-fallback SL multiplier = 1.5
• TP R-multiple (Full/Probation) = 2.0
• TP R-multiple (Trend tier) = 3.5 (wider — these strategies depend on tails)
• Agreement window (minutes) = 30

Bot enables — kill switches per bot. Useful for paper-A/B testing one allocation at a time.

Logging

Every entry prints:

[Ouroboros] OPEN Buy GER40 0.42 units @ 18243.50 | SL=18215.40 TP=18299.70 | tier=Full agree=2 risk=$15.00

Risk blocks (drawdown caps, max concurrent) and below-minimum-volume skips are also logged. Watch the cTrader Log tab.

How this was built

1. Six Python ML bots ran on six separate cTrader demo accounts for ~30 days, one timeframe each (M1→H4), collecting OHLCV bars and demo-traded signals.
2. PostgreSQL captured 644k signals and 3,253 closed trades.
3. Each cell of the (bot × symbol) matrix was scored: trade count, win-rate, total PnL, median PnL. Only cells that cleared thresholds were promoted.
4. Multi-bot agreement was tested as a gate (no effect on WR) and as a sizer (4-9× avg PnL uplift). Sizer won.
5. The four surviving Python strategies were ported to C# (cTrader's API), the whitelist was embedded as a compile-time table, and the cBot was wired with a SignalBus (rolling agreement) + PositionSizer (risk-balance×tier×agreement) + RiskGuard (Zero DD caps).

What's deliberately NOT in here

• No HTTP/network calls. Single .algo file, runs offline.
• No ML inference at runtime. All "ML" is baked into (a) the whitelist selection and (b) the strategy thresholds. Customers run pure C# logic.
• No anaconda H4. Data killed it.
• No taipan-USDJPY / hydra-USDCHF / viper-everywhere-but-BTC. Same.
• No hierarchical TF veto. Agreement is sizer-only.