Tradebot

An automated stock trading bot for an Alpaca paper trading account, plus a backtesting toolkit that tests whether its strategies are actually any good.

The bot runs every weekday after the US market close, checks the strategy you chose, and places a market order when the signal changes. The backtesters replay the same strategy code on years of daily and 5-minute history, so what you test is exactly what the bot trades.

The honest result: none of the 11 strategies tested (6 daily variants, 5 day-trading) beat simply buying and holding the same stock. The point of the project was to find that out with data instead of hope. See Results.

Features

• Live paper trading through the Alpaca API, with credentials kept in a .env file
• Five daily strategies in one interchangeable module: moving average crossover, monthly trend filter, 12-month momentum, RSI(2) dip buying and Donchian breakout
• Two trading modes: crossover (trade only on the day the signal flips) or follow (keep the account matched to the strategy, catching up after missed days)
• Safety checks: never sells shares it doesn't hold, never places the same order twice in one day, and has a dry-run mode
• Daily backtester with next-day-open fills, trading costs, buy-and-hold comparison, CAGR, max drawdown, Sharpe ratio and a full trade list
• Strategy comparison across symbols and periods in one command
• Intraday backtester for five published day-trading strategies on 5-minute bars, reporting month-by-month profit for a given budget in HUF
• Scheduled runs with macOS launchd, logging every run to bot.log
• Unit tests for the signal logic, strategies and simulators (offline, no API keys needed)

How it works

                 ┌─────────────────┐
  yfinance  ───► │ daily prices    │
                 └────────┬────────┘
                          ▼
                 ┌─────────────────┐     same code      ┌──────────────────┐
  config.py ───► │ strategies.py   │ ◄────────────────► │ backtest.py      │
                 │ position: 1 / 0 │                    │ history + stats  │
                 └────────┬────────┘                    └──────────────────┘
                          ▼
                 ┌─────────────────┐
                 │ bot.py          │  signal changed?  ──►  Alpaca market order
                 └─────────────────┘

Every strategy returns a daily position: 1 = hold the stock, 0 = stay in cash. Positions are decided at the close; the order fills at the next open. The bot and the backtester share this code, so the backtest reproduces the bot's behavior exactly.

Project structure

File	Purpose
bot.py	Live bot: fetch prices, run the strategy, place orders on Alpaca
config.py	Symbol, trade size, strategy, settings, mode, dry run
strategies.py	The five daily strategies
backtest.py	Daily backtester and multi-strategy comparison
intraday_strategies.py	Five day-trading strategies on 5-minute bars
intraday_backtest.py	Intraday backtester with monthly profit in HUF
scripts/make_charts.py	Regenerates the charts in docs/
deploy/com.mesziy.tradebot.plist	macOS launchd schedule (weekdays at 22:30 Budapest time)
tests/	Offline unit tests

Setup

Requires Python 3.10+ and a free Alpaca account with paper trading API keys.

git clone <this repo> tradebot && cd tradebot
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
cp .env.example .env        # then paste your Alpaca paper API key and secret

Check the connection (read-only, places no order):

.venv/bin/python -c "import bot; a = bot.get_client().get_account(); print(a.status, a.buying_power)"

Usage

Run the bot once:

.venv/bin/python bot.py

Run at 2026-10-25 11:22  |  strategy: ma_crossover {'short': 20, 'long': 50}
AAPL on 2026-10-23: close $347.73
  20d MA: 338.57 -> 338.91
  50d MA: 328.62 -> 329.47
  position: hold  |  mode: crossover
No signal today. No order placed.

Choose a strategy in config.py:

SYMBOL = "AAPL"
TRADE_QTY = 1
STRATEGY = "ma_crossover"            # trend_monthly, momentum_monthly, rsi2, donchian
STRATEGY_PARAMS = {"short": 20, "long": 50}
SIGNAL_MODE = "crossover"            # or "follow"
DRY_RUN = False

Backtest:

.venv/bin/python backtest.py                                        # config.py settings, 10 years
.venv/bin/python backtest.py --strategy momentum_monthly --symbol QQQ --mode follow --years 20
.venv/bin/python backtest.py --strategy ma_crossover --param short=50 --param long=200
.venv/bin/python backtest.py --compare --years 20                   # every strategy, 5 symbols
.venv/bin/python backtest.py --compare --years 20 --cash-rate tbill # idle cash earns T-bill interest
.venv/bin/python intraday_backtest.py --budget 20000 --symbols SPY QQQ

The intraday backtester downloads 5-minute bars from Alpaca on first use and caches them in data/.

Schedule it (macOS): edit the paths in deploy/com.mesziy.tradebot.plist, then:

mkdir -p ~/Library/LaunchAgents
cp deploy/com.mesziy.tradebot.plist ~/Library/LaunchAgents/
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.mesziy.tradebot.plist

Tests and charts:

.venv/bin/pip install -r requirements-dev.txt
.venv/bin/python -m pytest
.venv/bin/python scripts/make_charts.py

Results

All results use split- and dividend-adjusted prices, fills at the next day's open (or next 5-minute bar) and a trading cost per fill (0.05% daily, 0.03% intraday). Data runs to October 2026.

Idle cash earns 0% in the tables below unless stated, which slightly understates strategies that spend time out of the market. Cash interest shows the effect of adding it. The Sharpe ratio is computed without subtracting a risk-free rate.

The original bot: 20/50 crossover on AAPL, 10 years

	20/50 crossover	Buy & hold
Final value of $10,000	$42,704	$129,851
Yearly return (CAGR)	15.6%	29.2%
Biggest drop	−28.2%	−38.5%
Time invested	65%	100%
Trades	28 (50% winners)	1

The bot made money but captured only about half of the stock's yearly return. Late entries and whipsaw trades in sideways markets cost more than the crashes it avoided.

Five daily strategies, 20 years (yearly return / biggest drop)

Strategy	SPY	QQQ	AAPL
Buy & hold	11.0% / −55%	16.1% / −53%	28.3% / −61%
MA crossover 20/50	6.1% / −29%	7.8% / −31%	16.9% / −45%
MA crossover 50/200	8.8% / −34%	13.4% / −29%	20.8% / −59%
Trend 200-day, monthly (Faber 2007)	8.6% / −27%	12.5% / −29%	23.5% / −65%
Momentum 12-month, monthly (Moskowitz et al. 2012)	9.7% / −34%	15.1% / −29%	18.2% / −48%
RSI(2) dip buying (Connors 2008)	2.8% / −15%	3.1% / −14%	5.0% / −32%
Donchian 55/20 breakout	2.9% / −25%	6.9% / −20%	14.1% / −28%

Slow, monthly-checked trend and momentum rules on broad funds kept most of the return while roughly halving the 2008 crash. Faster rules gave up much more. None beat buy and hold.

Does interest on idle cash change anything?

A fair question from a reader: the strategies sit in cash roughly 10 to 90% of the time, and real cash earns interest. With --cash-rate tbill, idle cash earns the historical 13-week US T-bill yield (averaging 1.6% a year over these 20 years, 2.4% over the last 10).

Yearly return, 20 years	SPY: 0% cash → T-bills	QQQ	AAPL
Buy & hold	11.0%	16.1%	28.3%
MA crossover 20/50	6.1% → 6.5%	7.8% → 8.3%	16.9% → 17.5%
MA crossover 50/200	8.8% → 9.1%	13.4% → 14.0%	20.8% → 21.3%
Trend 200-day, monthly	8.6% → 8.9%	12.5% → 12.8%	23.5% → 23.9%
Momentum 12-month, monthly	9.7% → 10.0%	15.1% → 15.4%	18.2% → 18.4%
RSI(2) dip buying	2.8% → 4.2%	3.1% → 4.5%	5.0% → 6.4%
Donchian 55/20	2.9% → 3.7%	6.9% → 7.6%	14.1% → 14.9%

Interest adds 0.2 to 1.4 points a year, the most for strategies that are rarely invested. On the 10-year AAPL test it lifts the 20/50 bot from 15.6% to 16.7% a year, against 29.2% for buy and hold. It narrows the gap but changes no conclusion. The intraday backtester also leaves overnight cash uninvested.

Day trading with 20,000 HUF (about $61), 2019 to 2026

18 of 20 strategy/symbol combinations lost money after costs. Most had small positive returns before costs, but 10 to 80 trades a month at 0.03% each turned them negative. The one standout, a 30-minute range breakout on TSLA, earned most of its profit in 2019 to 2021 and became unprofitable at slightly higher trading costs.

What I learned

• Test the data, not the story. Every strategy here has a convincing rationale and published support. Most still lost to doing nothing.
• Trading costs decide intraday strategies. A tiny edge disappears after dozens of trades a month.
• Check how a backtest starts. A bot that only trades on signal flips can sit in cash for years if it starts mid-trend, which is why follow mode exists.
• One good result out of 20 is probably luck. Checking it year by year and at higher costs is what shows whether it holds up.

Disclaimer

This is an educational project that trades on a paper (simulated) account. It is not financial advice. Past performance does not predict future results.

License

MIT