🤖 DeepSeek Crypto Trading Bot

I built this trading bot to see if I could use AI (specifically DeepSeek LLM) to make trading decisions in the crypto market. Spoiler alert: it didn't work, but I learned a ton 📚

⚠️ Fair Warning

Don't use this with real money 💸❌ I tested three different strategies and they all lost money. This is a learning project, not a get-rich-quick scheme. The crypto market is way more complex than I initially thought.

💡 What This Project Is About

I wanted to experiment with combining AI language models and algorithmic trading. The idea was simple: fetch BTC price data from Bybit, ask DeepSeek to analyze it, and make buy/sell decisions based on its recommendations. Then backtest everything on historical data to see if it would've worked 🔍

📖 What I Actually Learned

• How to work with exchange APIs (Bybit in this case) 🔌
• Using LLMs for financial analysis (harder than it sounds) 🧠
• Building and running backtests 📊
• Working with financial data in Python 🐍
• Why most simple trading bots fail 💔

🛠️ Tech Stack

• Python 3.x 🐍
• Bybit API for market data 📈
• DeepSeek API for the AI part 🤖
• Pandas for data wrangling 🐼
• Basic HTTP requests 🌐

📁 Files in This Project

fetch_data.py          → Grabs BTC price history from Bybit 💰
deepseek_signal.py     → Asks DeepSeek for trading signals 🎯
backtest.py            → Tests the basic momentum strategy 📉
backtest_hybrid.py     → Tests technical indicators + AI combo 🔧
test_signal.py         → Quick tests to see what signals we get ⚡
btc_data.csv          → The historical data (auto-generated) 📄

🚀 How to Run This

Install Dependencies

pip install requests pandas

⚙️ Setup

1. You don't need Bybit API keys for fetching data (public endpoint) ✅
2. Get a DeepSeek API key from platform.deepseek.com 🔑
3. Drop your API key into the scripts where it says:

   DEEPSEEK_API_KEY = "your-api-key-here"

🏃 Running the Code

Fetch data:

python fetch_data.py

Test if signals are working:

python test_signal.py

Run the full backtest:

python backtest_hybrid.py

📊 The Strategies I Tried

Round 1: Simple Momentum 🎢

The idea: If price goes up more than 2.5% in 24 hours, buy. If it drops more than 2.5%, sell.

Result: Lost 1.15% over 1000 hours 📉 Made literally 1 trade.

Round 2: Let AI Do Everything 🤖

The idea: Give DeepSeek all the price data and let it figure out patterns.

Result: Lost 0.65% 📉 Still only 1 trade. AI was way too cautious.

Round 3: Hybrid Approach ⚡

The idea: Use actual technical indicators (RSI, moving averages, volume) and have DeepSeek confirm the signals.

Result: Lost 3.95% 💀 Worst performance yet. 0% win rate.

🤔 Why Everything Failed

The Main Issues

1. DeepSeek was too conservative 🛑 - It basically said "HOLD" for everything. Even when I gave it clear rules, it barely pulled the trigger.

2. Crypto is chaotic 🌪️ - BTC dropped 3% over the test period and the bot just couldn't adapt fast enough.

3. My prompts weren't good enough 📝 - Getting an LLM to consistently make good trading decisions is really hard.

4. Not enough trades ⏸️ - 1-2 trades over 1000 hours means the strategy is way too picky.

💭 What I Realized

Real trading bots that actually make money need:

• Way more sophisticated algorithms 🧮
• High-frequency trading capabilities ⚡
• Proper risk management 🛡️
• Years of development and testing ⏳
• A lot more computing power than my laptop 💻

LLMs are probably better for:

• Analyzing news sentiment 📰
• Summarizing market reports 📋
• Supporting decisions, not making them 🤝

And the biggest lesson: always backtest before risking real money 💯 Even then, past performance doesn't mean future success.

🔮 If I Were to Continue This

Some ideas I didn't try:

• Copy trading 👥 - Just follow people who actually know what they're doing
• Grid bots 🎯 - Buy low, sell high in a price range (works in sideways markets)
• DCA bots 💰 - Invest small amounts regularly over time
• Sentiment analysis 📊 - Use AI to read crypto Twitter instead of predicting prices
• Multiple LLMs voting 🗳️ - Get several AI opinions and average them out

🤝 Want to Use This?

Feel free to fork it and experiment 🍴 Just remember this didn't make money, so don't expect miracles. It's a learning project, not a money printer 🖨️❌

📄 License

MIT - Do whatever you want with this code ✅

🙏 Thanks

• Bybit for free API access 🎁
• DeepSeek for the LLM API 🤖
• Everyone who told me this wouldn't work (you were right) 😅

---

Bottom line: I set out to build a trading bot and ended up learning why that's really hard 🎓 The code works, the backtesting works, but the strategy doesn't. That's still valuable experience though - now I know what doesn't work and why 💡