Aster Timesfm Pipeline

Aster → TimesFM Forecasting Pipeline

Overview

End-to-end crypto price forecasting: pull historical klines from Aster DEX via MCP, transform to numpy arrays, run through TimesFM 2.5 for zero-shot point forecasts with quantile prediction intervals.

When to Use

• Forecasting crypto pair prices (BTC/USDT, ETH/USDT, SOL/USDT, etc.)
• Generating trading signals with confidence intervals
• Risk assessment via quantile spreads (10th-90th percentile)
• Comparing forecast vs actual for strategy validation

Pipeline Steps

Step 1: Pull Market Data via Aster MCP

# Use mcp__aster__get_klines to pull historical data
# Example: 500 hourly candles for BTCUSDT
klines = mcp__aster__get_klines(symbol="BTCUSDT", interval="1h", limit=500)

Available intervals: 1m, 5m, 15m, 1h, 4h, 1d

Step 2: Transform to TimesFM Input

import numpy as np

# Extract close prices from klines response
# Klines format: [open_time, open, high, low, close, volume, ...]
close_prices = np.array([float(k[4]) for k in klines])
volumes = np.array([float(k[5]) for k in klines])

# Optional: use funding rates as covariate
# funding = mcp__aster__get_funding_rate(symbol="BTCUSDT")

Step 3: Run TimesFM Forecast

import timesfm

# Load model (cached after first call, ~800MB download)
tfm = timesfm.TimesFM_2p5_200M_torch()

# Forecast next N periods
horizon = 24  # 24 hours ahead for 1h candles
point_forecast, quantile_forecast = tfm.forecast(
    horizon=horizon,
    inputs=[close_prices],
)

# point_forecast shape: (1, horizon) — predicted prices
# quantile_forecast shape: (1, horizon, num_quantiles) — prediction intervals

Step 4: Interpret Results

# point_forecast[0] = array of predicted prices for next 24 hours
# quantile_forecast[0] = array of [10th, 20th, ..., 90th] percentile predictions

current_price = close_prices[-1]
predicted_price = point_forecast[0][-1]  # price at end of horizon
change_pct = (predicted_price - current_price) / current_price * 100

q10 = quantile_forecast[0][-1][0]  # 10th percentile (downside)
q90 = quantile_forecast[0][-1][-1]  # 90th percentile (upside)

print(f"Current: {current_price:.2f}")
print(f"Forecast ({horizon}h): {predicted_price:.2f} ({change_pct:+.2f}%)")
print(f"Range: {q10:.2f} to {q90:.2f}")

Step 5: Multi-Pair Batch Forecasting

pairs = ["BTCUSDT", "ETHUSDT", "SOLUSDT"]
all_series = []

for pair in pairs:
    klines = mcp__aster__get_klines(symbol=pair, interval="1h", limit=500)
    closes = np.array([float(k[4]) for k in klines])
    all_series.append(closes)

# Batch forecast all pairs at once (much faster than one-by-one)
point_forecasts, quantile_forecasts = tfm.forecast(
    horizon=24,
    inputs=all_series,
)

Risk Rules

1. Never trade solely on forecast — TimesFM is general-purpose, not crypto-specific
2. Validate accuracy first — backtest on historical data before live trading
3. Wide quantile spread = high uncertainty — reduce position size when q90-q10 is large
4. Short timeframes are noisier — 1h/4h more reliable than 1m/5m for crypto
5. Model freshness — re-pull klines before each forecast, stale data = stale predictions

Integration with Other Skills

• anomaly-detector: Alert when actual price breaks outside forecast quantile bounds
• autoresearch: Measure forecast accuracy over time, optimize horizon/interval parameters
• council: Present forecasts to multi-expert panel for trading decisions
• MiroFish: Combine quantitative TimesFM forecast with qualitative scenario simulation

Environment

# TimesFM venv
source ~/Projects/timesfm/.venv/bin/activate

# Required: timesfm>=2.0.0, torch, numpy
# Model: google/timesfm-2.5-200m-pytorch (auto-downloads from HuggingFace)