An autonomous Bitcoin trading bot with deep-learning price prediction — built in 2018.
FleckBot watched the Bitfinex BTC/USD market and put a neural network behind the trading decision: LSTM networks (DeepLearning4j) trained on OHLCV history to predict the next period's price, a prediction-driven strategy backtested against a dual-SMA baseline on four years of Coinbase minute data, and a Spring MVC dashboard (in Portuguese) to drive the whole thing.
2026 — FleckBot is being reborn. Same bot, new brain: the LSTM has been replaced by Claude as the market analyst, with a human approving every trade and an honest scoreboard measuring whether the bot actually beats buy & hold. The original interface you see here is being faithfully recreated in React as part of the rebuild.
| Page | Feature |
|---|---|
| Análise | Live BTC/USD technical-analysis chart (TradingView) |
| Carteiras | Exchange wallet balances across 10 assets, with allocation pie chart |
| Playground | Backtesting sandbox: dual-SMA or LSTM-prediction strategies over 2014–2018 Coinbase minute data, any timeframe from 1 minute to 1 month — orders, fees, P&L vs HODL |
| Redes Neurais | Train LSTM networks on any date range and price field (close/open/high/low/volume), then chart predicted vs. actual price with error metrics |
| Logs | Paginated engine activity log |
- Trading engine — a single-threaded control loop (inspired by BX-bot) that loads exchange, market, and strategy configuration from an XML datastore, monitors an emergency-stop balance, and emails a critical alert if anything goes wrong.
- Exchange adapters — Bitfinex API v1 (HMAC-SHA384 signed) for live data and orders, plus a dry-run test adapter backed by Bitstamp public data.
- Neural nets — 2× GravesLSTM (256 units) → dense → regression head (DL4J), min-max-normalized 5-feature OHLCV input, 178-bar window, one-step-ahead prediction, MSE loss.
- Backtesting — ta4j strategies over Coinbase 1-minute bars downsampled to the chosen timeframe; per-trade fee accounting (0.1%/side) and buy-&-hold comparison.
Java 8 · Spring MVC 4.3 · Spring Security · DeepLearning4j 0.9 · ta4j · JSP/Bootstrap 4 · PostgreSQL · JAXB XML config store
- Database — edit
src/main/resources/application.propertieswith your PostgreSQL url/login/password. - Exchange — add your own API credentials to
src/main/resources/xml/exchange.xml(never commit real keys). - Historical data — the Playground and neural-net pages expect the Kaggle Coinbase dataset (
coinbaseUSD_1-min_data_2014-12-01_to_2018-01-08.csv) in~/csv/. - Login — development credentials are defined in
conf/SecurityConfig.java.
This repo is kept as-is as the origin of the FleckBot story. Active development happens in the V2 rebuild.