Twap detection¶

Detect simple TWAP executions in tick trades data.

We use crypto-lake.com sample/free market data, FTRB-USDT market on Ascendex.

Quick links:

edit this notebook online using Binder
follow our activity on twitter

In [1]:

import datetime
import math

import numpy as np
import pandas as pd
import cufflinks as cf
from scipy import signal
import plotly.express as px

import lakeapi

cf.go_offline()
lakeapi.use_sample_data(anonymous_access=True)

In [2]:

# Parameters
symbol = 'FTRB-USDT'
exchange = 'ASCENDEX'

# Free sample data contain subset of the below time period
start = datetime.datetime(2022, 9, 1)
end = datetime.datetime(2022, 12, 15)

Data¶

In [11]:

print('Loading trades')
trades = lakeapi.load_data(
    table = 'trades',
    start = start,
    end = end,
    symbols = [symbol],
    exchanges = [exchange],
    drop_partition_cols = True,
).sort_values('origin_time')

Loading trades

  0%|          | 0/81 [00:00<?, ?it/s]

Add features¶

In [13]:

trades['signed_quantity'] = trades.quantity * trades.side.map({'buy': +1, 'sell': -1}).astype(int)

# A trick here: instead of using trade side as 'aggressor', I use something else as aggressivity: 
# direction and strength of current short-term trend.
trades['price_ewm'] = trades.price.ewm(halflife = 5).mean()
trades['aggressivity'] = (trades.price - trades.price_ewm) / trades.price_ewm
trades['aggr_quantity'] = trades.quantity * trades.aggressivity

print(trades.shape)
trades[100:].head(2)

(99539, 10)

Out[13]:

	side	quantity	price	trade_id	origin_time	received_time	signed_quantity	price_ewm	aggressivity	aggr_quantity
100	sell	2586.0	0.010007	108086470187312032	2022-09-01 15:57:58.421999872	2022-09-01 15:57:58.613365504	-2586.0	0.010023	-0.001642	-4.246117
101	buy	2550.0	0.010013	108086470187322384	2022-09-01 15:58:23.803000064	2022-09-01 15:58:24.033291776	2550.0	0.010022	-0.000908	-2.316488

In [14]:

def volume_share_plot(selector, title = None, resample = '60min'):
	return (
		trades[selector].set_index('origin_time').resample(resample).quantity.sum() 
		/ 
		trades.set_index('origin_time').resample(resample).quantity.sum() 
	).iplot(title = title, yTitle = 'volume share')

def volume_plot(selector, title = None, resample = '60min'):
	df = trades
	return df[selector].set_index('origin_time').resample(resample).agg({'aggr_quantity': 'sum'}).iplot(title = title, yTitle = 'adjusted imbalance')

Frequency analysis¶

In [15]:

# Estimate power spectral density using a periodogram
f, Pxx = signal.periodogram(trades.quantity, fs = 1e6, nfft = 1000)
plot = px.line(x = f, y = Pxx, title = 'Peak frequencies (in ms)')
for i in [6, 60, 5*60]:
	plot.add_vline(i * 1_000, line_width=1, line_color="orange")
plot

We see a lot of quantity is submited in frequencies of 6 seconds, 60 seconds and 5 minutes. Lets focus on 60-second twap orders aligned to second boundary = eg. around 12:35:45.000 +- a few millis.

Minute twap¶

In [16]:

around_whole_second = (trades.origin_time.dt.microsecond < 80_000) | (trades.origin_time.dt.microsecond > 950_000)

volume_share_plot(selector = around_whole_second, resample = '60min')
volume_plot(selector = around_whole_second, resample = '60min')
trades.set_index('origin_time').price[::100].iplot(yTitle = 'price')

We can see a few strong twaps in the charts. Eg. Nov 7 -- Nov 11 had a significant sell twap, it's effect is clearly visible in price chart as well. Other sells are around Dec 4 and perhaps Nov 2. One buy twap with smaller volume is around Oct 25. Some of the twap chunks don't seem to be imbalanced and therefore affecting price, eg. in September.

Conclusion¶

We detected a few twap executions, some of which were affecting price significantly. This analysis can be extended to run in real-time to serve as a trading signal or MM bot adjustment. The advantage of filtering out twap data compared to following price trend is that the signal is more clearly readable and can be easily used for trend-following.