Deep Temporal Convolutional Networks for High-Frequency Cryptocurrency Price Forecasting
Log in to collectOnsite backtest IDE
Quant Buffet native backtest IDEEdit and run Quant Buffet Python for Deep Temporal Convolutional Networks for High-Frequency Cryptocurrency Price Forecasting in the browser. Results update live with equity, drawdown, and metrics charts. Allowed: backtest.data, backtest.engine, backtest.metrics, numpy, pandas. Define ASSETS and make_on_day(prices). Shortcut: Ctrl+Enter. API docs →
Quant Buffet syntax cheat sheet (copy / insert)
Paste these fragments into the editor. The sandbox rejects QuantConnect, os, and network libraries.
from __future__ import annotations
import numpy as np
import pandas as pd
from backtest.data import load_daily_prices
from backtest.engine import EngineConfig, PortfolioEngine
from backtest.metrics import compute_metricsASSETS = ["SPY", "QQQ", "TLT", "GLD", "BIL"]def make_on_day(prices: pd.DataFrame):
cols = [c for c in ASSETS if c in prices.columns]
sma = prices[cols].rolling(200, min_periods=200).mean()
state = {"last": None}
def on_day(engine: PortfolioEngine, dt: pd.Timestamp) -> None:
if sma.loc[dt].isna().all():
return
key = (dt.year, dt.month)
if state["last"] == key:
return
state["last"] = key
long = [
s for s in cols
if pd.notna(prices.at[dt, s]) and pd.notna(sma.at[dt, s])
and prices.at[dt, s] > sma.at[dt, s]
]
weights = {} if not long else {s: 1.0 / len(long) for s in long}
engine.set_target_weights(dt, weights)
ready = sma.dropna(how="all").index.min() if sma.notna().any().any() else None
return on_day, readyengine.set_target_weights(dt, {"SPY": 0.60, "BIL": 0.40})Live backtest performance
Accent = strategy · dashed grey = buy-and-hold benchmark
Export to your platform
Transform Quant Buffet lab code (ASSETS + make_on_day / PortfolioEngine) into native classes for a third-party IDE — then copy and paste.
# Generated from Quant Buffet → QuantConnect LEAN
# Strategy: Deep Temporal Convolutional Networks for High-Frequency Cryptocurrency Price Forecasting
# Detected pattern: Mean reversion
# Source uses Quant Buffet lab APIs (ASSETS + make_on_day / PortfolioEngine).
# Review fees, data, and risk before live trading — educational export only.
from AlgorithmImports import *
class QuantBuffetExport(QCAlgorithm):
def Initialize(self):
self.SetStartDate(2010, 1, 1)
self.SetCash(100000)
tickers = ["SPY", "TLT", "GLD", "BIL"]
self.symbols = []
for t in tickers:
if "-" in t: # crypto proxy e.g. BTC-USD
self.symbols.append(self.AddCrypto(t.replace("-USD", ""), Resolution.Daily).Symbol)
else:
self.symbols.append(self.AddEquity(t, Resolution.Daily).Symbol)
self.Schedule.On(
self.DateRules.MonthStart(self.symbols[0]),
self.TimeRules.AfterMarketOpen(self.symbols[0], 30),
self.Rebalance,
)
# Logic: Buy when return z-score < -1 over 20 days.
def Rebalance(self):
import numpy as np
picks = []
for symbol in self.symbols:
hist = self.History(symbol, 20 + 5, Resolution.Daily)
if hist.empty: continue
close = hist["close"]
if hasattr(close, "unstack"):
close = close.unstack(level=0).iloc[:, 0]
rets = close.pct_change().dropna()
if len(rets) < 20: continue
window = rets.iloc[-20:]
z = (window.iloc[-1] - window.mean()) / (window.std() or 1e-9)
if z < -1:
picks.append(symbol)
w = 1.0 / len(picks) if picks else 0.0
for symbol in self.symbols:
self.SetHoldings(symbol, w if symbol in picks else 0.0)
Exported code uses the platform’s native classes and libraries. Install dependencies in your third-party IDE, then run. Validate before live trading.
Academic paper
Deep Temporal Convolutional Networks for High-Frequency Cryptocurrency Price Forecasting
Yizhou Jin; Boyi Huang; Zaixiao Peng; Xu Miao
https://www.semanticscholar.org/paper/04b0bd378e3d97b43d6bdbc85aec2dbe5dc9e2f6
Teaser
Enter when short-horizon z-score is deeply negative; exit near zero. Universe: BTC-USD, ETH-USD. Parameters: lookback=20; entry_z=-1.0; exit_z=0.0; rebalance=daily. Rebalanced on the engine's template schedule with 5 bps commission and 2 bps slippage. Because the paper's primary signal (ML, sentiment, or proprietary data) is not available in our public ETF engine, this draft uses a liquid ETF rule that preserves the paper's economic theme rather than a bit-exact replication.
Strategy in a nutshell
High-frequency cryptocurrency price prediction remains challenging due to extreme volatility, market noise, and complex temporal dependencies. Traditional machine learning methods, recurrent neural networks (RNNs), and attention-enhanced models have shown limited predictive accuracy and generalization on such data. To address these limitations, this study systematically evaluates Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Temporal Convolutional Networks (TCNs) using multi-factor inputs. The dataset spans 2018–2024 and includes OHLC prices, trading volume, number of trades, short-term buying pressure, and technical indicators such as moving averages, EMA deviations, MACD, Bollinger Bands, and momentum. Models were trained using standard proc
Economic rationale
Short-horizon overreaction produces temporary dislocations that reverse toward a local mean. Related evidence from “Deep Temporal Convolutional Networks for High-Frequency Cryptocurrency Price Forecasting”: High-frequency cryptocurrency price prediction remains challenging due to extreme volatility, market noise, and complex temporal dependencies. Traditional machine learning methods, recurrent neural networks (RNNs), and attention-enhanced models have shown limited predictive accuracy and generalization on such data. To address these limitations, this study systematically evaluates Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Temporal Convolutional N