Concurrent Scalping Algo
This python script is a working example to execute scalping trading algorithm for Alpaca API. This algorithm uses real time order updates as well as minute level bar streaming from Polygon via Websockets (see the document for Polygon data access). One of the contributions of this example is to demonstrate how to handle multiple stocks concurrently as independent routine using Python's asyncio.
The strategy holds positions for very short period and exits positions quickly, so you have to have more than $25k equity in your account due to the Pattern Day Trader rule, to run this example. For more information about PDT rule, please read the document.
Dependency
This script needs latest Alpaca Python SDK. Please install it using pip
$ pip3 install alpaca-trade-api
or use pipenv using Pipfile
in this directory.
$ pipenv install
Usage
$ python main.py --lot=2000 TSLA FB AAPL
You can specify as many symbols as you want. The script is designed to kick off while market is open. Nothing would happen until 21 minutes from the market open as it relies on the simple moving average as the buy signal.
Strategy
The algorithm idea is to buy the stock upon the buy signal (20 minute
moving average crossover)
as much as lot
amount of dollar, then immediately sell the position at or above the entry price.
The assumption is that the market is bouncing upward when this signal occurs in a short period of time.
The buy signal is extremely simple, but what this strategy achieves is the quick reaction to
exit the position as soon as the buy order fills. There are reasonable probabilities that you can sell
the positions at the better prices than or the same price as your entry within the small window. We send
limit order at the last trade or position entry price whichever the higher to avoid unnecessary slippage.
The buy order is canceled after 2 minutes if it does not fill, assuming the signal is not effective anymore. This could happen in a fast-moving market situation. The sell order is left indifinitely until it fills, but this may cause loss more than the accumulated profit depending on the market situation. This is where you can improve the risk control beyond this example.
The buy signal is calculated as soon as a minute bar arrives, which typically happen about 4 seconds after the top of every minute (this is Polygon's behavior for minute bar streaming).
This example liquidates all watching positions with market order at the end of market hours (03:55pm ET).
Implementation
This example heavily relies on Python's asyncio. Although the thread is single, we handle multiple symbols concurrently using this async loop.
We keep track of each symbol state in a separate ScalpAlgo
class instance. That way,
everything stays simple without complex data structure and easy to read. The main()
function creates the algo instance for each symbol and creates streaming object
to listen the bar events. As soon as we receive a minute bar, we invoke event handler
for each symbol.
The main routine also starts a period check routine to do some work in background every 30 seconds. In this background task, we check market state with the clock API and liquidate positions before the market closes.
Algo Instance and State Management
Each algo instance initializes its state by fetching day's bar data so far and position/order from Alpaca API to synchronize, in case the script restarts after some trades. There are four internal states and transitions as events happen.
TO_BUY
: no position, no order. Can transition toBUY_SUBMITTED
BUY_SUBMITTED
: buy order has been submitted. Can transition toTO_BUY
orTO_SELL
TO_SELL
: buy is filled and holding position. Can transition toSELL_SUBMITTED
SELL_SUBMITTED
: sell order has been submitted. Can transition toTO_SELL
orTO_BUY
Event Handlers
on_bar()
is an event handler for the bar data. Here we calculate signal that triggers
a buy order in the TO_BUY
state. Once order is submitted, it goes to the BUY_SUBMITTED
state.
If order is filled, on_order_update()
handler is called with event=fill
. The state
transitions to TO_SELL
and immediately submits a sell order, to transition to the
SELL_SUBMITTED
state.
Orders may be canceled or rejected (caused by this script or you manually cancel them
from the dashboard). In these cases, the state transitions to TO_BUY
(if not holding
a position) or TO_SELL
(if holding a position) and wait for the next events.
checkup()
method is the background periodic job to check several conditions, where
we cancel open orders and sends market sell order if there is an open position.
It exits once the market closes.
Note
Each algo instance owns its child logger, prefixed by the symbol name. The console
log is also emitted to a file console.log
under the same directory for your later review.
Again, the beautify of this code is that there is no multithread code but each algo instance can focus on the bar/order/position data only for its own. It still handles multiple symbols concurrently plus runs background periodic job in the same async loop.
The trick to run additional async routine is as follows.
loop = stream.loop
loop.run_until_complete(asyncio.gather(
stream.subscribe(channels),
periodic(),
))
loop.close()
We use asyncio.gather()
to run all bar handler, order update handler and periodic job
in one async loop indifinitely. You can kill it by Ctrl+C
.
Customization
Instead of using this buy signal of 20 minute simple moving average cross over, you can
use your own buy signal. To do so, extend the ScalpAlgo
class and write your own
_calc_buy_signal()
method.
class MyScalpAlgo(ScalpAlgo):
def _calculate_buy_signal(self):
'''self._bars has all minute bars in the session so far. Return True to
trigger buy order'''
pass
And use it instead of the original class.