Closed beta · 500M+ timesteps trained

A deep RL agent that learns to grow your book
through millions of simulated markets.

Most algorithmic systems run on static rules — buy when RSI crosses 30, sell at resistance. They don't know NFP is in two hours, or that the Fed just hiked. Piggy is a neural network that fuses 138M+ price bars with the live economic calendar and sentiment from 80+ news sources — and adapts its behaviour to the regime it's actually in.

Request early access See how it learns →

market features per bar

timeframes, 1m to 1d

80+

live news sources

138M+

price bars trained on

24/7

Gemma 4 on GPU

What Piggy actually does

73 technical features × 7 timeframes. One coherent view of the market.

Every decision considers the full picture — 1m, 5m, 15m, 30m, 1h, 4h and 1d — across 73 engineered features grouped into nine analytical categories. The agent learns which signals matter in which conditions, instead of trusting any single one in isolation.

📊

Price action

Returns, range, body ratio, wick lengths, gaps and inside bars — the raw geometry of each candle, normalised across instruments.

📈

Trend

SMA 10 / 20 / 50, EMA crossovers and market-structure analysis — the directional bias every desk trades around.

⚡

Momentum

RSI, MACD, Williams %R, CCI and ROC — the full momentum suite, computed on every timeframe Piggy watches.

🌡️

Volatility

Bollinger Bands, ATR and ADX with +DI / −DI — the vol structure that drives position sizing and stop placement.

☁️

Ichimoku cloud

Full Ichimoku: Tenkan, Kijun, Senkou A and Senkou B — the Japanese system that captures trend, momentum and equilibrium together.

📐

Levels

Fibonacci retracements, pivot points and support / resistance detection — the geometry of where price actually pauses.

🕯️

Candlestick patterns

Doji, hammer, engulfing, morning and evening star — recognised in context, not treated as standalone signals.

🔀

Divergences

RSI and MACD bullish / bearish divergences — the early warning when momentum and price stop agreeing.

🕒

Context

Session times, market regimes and cross-pair correlation — so the agent reads the same chart differently in different worlds.

News intelligence · continuous

80+ news sources. Read every second.
Zero sleep.

A market doesn't wait for the open, and neither does Piggy. A dedicated language-model pipeline reads more than eighty news sources every second, tracking sentiment shifts across FX, equities, futures, crypto and the wider commodities complex — and feeds those shifts straight into the trading policy as a first-class input.

✓80+ sources polled continuously, every second of every day.
✓Multi-asset coverage — FX, equities, futures, crypto and commodities, each headline tagged to the instruments it actually moves.
✓Gemma 4 on state-of-the-art GPUs — production-grade LLM inference, 24/7, no batch lag.
✓Sentiment deltas in the policy — change-in-tone is an input the agent trades on, not a dashboard you read after the fact.
✓Continuous training — the same GPU fleet retrains the agent on fresh data alongside live inference, so the model never goes stale.

News pipeline · live

80+ sources active · 0 ms idleGemma 4 · read + train, 24/7

Stocks

Futures

Crypto

Commod

EURUSD · sentiment Δ

+0.42

BTCUSD · sentiment Δ

+0.18

XAUUSD · sentiment Δ

−0.27

Economic calendar · integrated

It knows NFP is in two hours. It trades like it.

Static bots blunder into rate decisions and payrolls prints because they don't read the calendar. Piggy tracks every high-impact macro event in real time and adjusts risk, stops and entries around them.

🏛️

Central bank decisions

FOMC · ECB · BOE · BOJ · RBA · BOC · SNB · RBNZ — every G10 rate decision, statement and press conference, tagged to the pairs they actually move.

📋

Employment & inflation

US Non-Farm Payrolls and unemployment, CPI & Core CPI for the US, UK and EU, PPI, PCE price index and initial jobless claims.

📊

Growth & sentiment

GDP releases, retail sales, PMI (manufacturing & services), consumer confidence and German ZEW sentiment.

How Piggy uses it

✓Reduce position sizes before high-impact events
✓Widen stops during volatile releases
✓Avoid new entries within 2 hours of major news
✓Track surprise magnitude as a momentum signal

Risk built into the brain

Risk isn't a bolt-on. It's part of the policy.

Most systems treat risk as a wrapper around a signal — a stop-loss on top of a buy. In Piggy, position sizing, stops and targets are decisions the policy learned end-to-end alongside the entry. The agent was rewarded for surviving as much as for winning.

✓1% risk per trade — ATR-based position sizing, scaled by current volatility on every entry.
✓Stop loss at 2× ATR — set from entry, adapted per instrument and regime.
✓Take profit at 3× ATR — a 1.5 : 1 reward-to-risk floor on every trade.
✓Max 4 simultaneous positions — capped at 80% total exposure, so the book is never all-in.
✓Notional sizing with 10% margin — leverage handled properly for FX and crypto, not assumed away.
✓Drawdown penalty in training — the reward function punishes equity destruction, so the agent learns to avoid it.

Sample trade · EURUSD · 1H

Long 0.42 lots · risk 1.0% of equityATR(14) = 11.4 pips · regime: trending

Entry

Size

Stop

Target

Exit

Stop (2× ATR)

−23p

Target (3× ATR)

+34p

R : R

1.5×

How it learns

Trained with PPO on 138 million price bars.

Piggy is trained using Proximal Policy Optimisation — the same family of reinforcement-learning algorithms behind state-of-the-art game-playing and robotics agents — over 500 million+ timesteps on real market data, with 2,048 parallel environments running in lockstep. Thousands of years of trading experience, compressed into GPU compute.

📊

Sharpe-first reward

The agent is optimised for risk-adjusted return, not raw profit. A jumpy equity curve gets less reward than a steady one with the same end value.

⚖️

Asymmetric trade quality

Losing trades are penalised more than equivalent wins are rewarded, so the policy learns to avoid bad setups rather than chase good-looking ones.

🛡️

Capital preservation

Drawdown beyond defined thresholds incurs heavy penalty during training — the agent treats large losses as catastrophic, not as a cost of doing business.

Target performance

The bars Piggy has to clear.

These are the acceptance criteria a Piggy model has to hit, on out-of-sample data and across multiple market regimes, before any capital — paper or real — goes near it.

> 1.5

Sharpe ratio

< 15%

Max drawdown

> 50%

Win rate

+EV

across all regimes

walk-forward folds positive

Targets reflect Piggy's internal acceptance criteria. Past simulated performance is not a guarantee of future results. Trading involves risk of loss.

Markets covered

FX first. Crypto next.

Piggy is trained today on the deepest, longest-history liquid markets in the world. Crypto coverage rolls in next, with the same reinforcement-learning core.

💱

FX majors

EUR/USD, GBP/USD, USD/JPY, USD/CHF, AUD/USD, USD/CAD, NZD/USD — every G10 pair where liquidity and history are deepest.

🌍

FX crosses

EUR/GBP, GBP/JPY, EUR/JPY, AUD/JPY and CAD/JPY — where regime context and cross-pair correlation matter most.

🪙

Crypto · coming

BTC, ETH and liquid alts — same architecture, retrained on crypto microstructure and 24/7 session dynamics.

Technology stack

Built like a real trading system, not a notebook.

Production-grade plumbing under the model — the same kind of stack a quant fund would build for a desk of PMs, just running for one agent instead.

🧠

Learning

Proximal Policy Optimisation (PPO) — a stable, modern reinforcement-learning algorithm proven across robotics and game-playing.

⚙️

Execution

cTrader Open API for FX, Binance for crypto — direct broker integration, no third-party signal layer in the middle.

🗄️

Data

138M+ historical price bars, the live economic calendar and 80+ news sources, fused in one observation pipeline.

⚡

Infrastructure

GPU-accelerated training with 2,048 parallel environments — the agent simulates thousands of market years per training day.

Why reinforcement learning?

Traditional ML predicts direction.
RL learns the whole strategy.

Supervised models answer one narrow question — "is the next bar up or down?" — and leave entries, sizing, exits and risk management to whatever rules you bolt on around them. The model is optimised for prediction accuracy, not for the thing that actually matters: how the equity curve looks at the end of the year.

Reinforcement learning is different. The agent learns a complete policy — when to enter, how much to risk, when to exit, how to manage multiple positions — directly from the trading outcomes it produces. Every decision is optimised against the same objective the trader cares about.

Supervised ML vs reinforcement learning

Optimised for the right thingtrading outcomes, not prediction accuracy

Direction predict

Entry timing

Position sizing

Stop / target

Portfolio mgmt

Who it's for

Built for people who already know why this is hard.

🎯

Serious traders

You want systematic, emotion-free execution on a book you already understand — not a copy-trading widget for someone else's calls.

🏦

Fund managers

Emerging or established — you want multi-strategy diversification under one adaptive system instead of staffing six PMs to cover the same ground.

🔬

Quant researchers

You want a reinforcement-learning testbed with real market data, real broker plumbing and a risk layer that isn't an afterthought.

🐷

Long-term compounders

You optimise for risk-adjusted returns over years, not screenshots. Steady Sharpe beats a flashy month.

Beta · iPhone & Android

Take Piggy with you.

Open the live beta in Expo Go — watchlist, live trade tickets, positions, orders and portfolio, all secured behind biometric unlock. Identical on iPhone and Android.

Scan in Expo Go

1 · Install Expo Go (App Store / Play Store).
2 · Scan this code — iPhone users can use the Camera app too.

Prefer a standalone Android build?

Download Android APK

Piggy doesn't predict the market. It learns to survive it.