Crypto Doesn't Exist in a Vacuum
Crypto traders tend to focus on charts — RSI, MACD, support and resistance. But crypto prices are influenced by forces far beyond the charts.
When the Federal Reserve raises interest rates, crypto sells off. When the US dollar strengthens, crypto weakens. When the VIX spikes, risk assets across the board get hit.
We tested whether adding these macro-economic indicators to our prediction model would improve accuracy. The answer: yes, by 3-5%.
The Macro Indicators We Tested
We pull daily data from public economic data APIs and align it with hourly crypto candles:
1. US Treasury Yields
10-Year Yield (DGS10): The benchmark "risk-free" rate. When yields rise, investors can earn more from bonds — reducing the appeal of risky assets like crypto.
2-Year Yield (DGS2): Reflects short-term interest rate expectations. Closely tied to Federal Reserve policy.
Yield Curve (10Y-2Y spread): The difference between long and short rates. When this inverts (goes negative), it historically signals economic slowdown. Crypto reacts to these macro shifts.
2. VIX (Volatility Index)
The VIX measures expected volatility in the US equity market. It's often called the "fear gauge."
- VIX spikes (>25) = fear = risk-off = crypto sells
- VIX low (<15) = complacency = risk-on = crypto rallies
- VIX rate of change is even more predictive than the absolute level
3. US Dollar Index (DXY)
The dollar and crypto have an inverse relationship. When the dollar strengthens:
- Capital flows out of risk assets into dollar-denominated safe havens
- Emerging market liquidity tightens
- Crypto typically weakens
The 20-day rate of change in the dollar index was one of our most predictive macro features.
4. S&P 500
Crypto has become increasingly correlated with equities since 2020. The S&P 500's momentum (especially its 20-day rate of change) provides useful context for crypto direction.
5. Gold
Gold and crypto sometimes move together (both "alternative" stores of value) and sometimes diverge (gold as safe haven during crypto selloffs). The relationship is complex but informative.
Feature Engineering: 28 Macro Features
We don't just feed raw prices into the model. From the 7 base indicators, we derive 28 features:
- Raw values (normalised): yield_10y, yield_2y, yield_curve, vix, sp500, dxy, gold
- 5-day rate of change for each indicator (captures short-term momentum)
- 20-day rate of change for each (captures medium-term trends)
- Central bank proximity: Days until next Federal Reserve meeting, days since last meeting, and a binary "FOMC week" flag
The FOMC proximity feature is interesting — markets tend to become more volatile and directional in the days surrounding rate decisions.
The Experiment
Test K3 in our experiment series. We compared four configurations:
- Chart-only baseline: 60+ technical features, no macro data
- Macro-only: Only the 28 macro features, no chart data
- Chart + Macro: Both feature sets combined
- Chart + Macro + Lag/ROC: Plus lagged values and rate-of-change of chart indicators
All tests used XGBoost Aggressive (200 trees, depth 8) with walk-forward validation across 10 cryptocurrencies.
Results
| Configuration | Average Accuracy | Improvement |
|---|---|---|
| Chart-only baseline | 51.8% | — |
| Macro-only | 55.0% | +3.2% |
| Chart + Macro | 56.1% | +4.3% |
| Chart + Macro + Lag/ROC | 56.3% | +4.5% |
Macro features alone outperformed chart features alone. That's a striking result. It suggests that for hourly crypto prediction over a 20-candle horizon, macro context matters more than traditional technical indicators.
The combined approach was best, as expected — chart features capture short-term price dynamics while macro features capture the broader environment.
Per-Coin Breakdown
Not all coins responded equally to macro features:
| Coin | Chart Only | Chart+Macro | Improvement |
|---|---|---|---|
| BTC | 51.0% | 59.2% | +8.2% |
| ETH | 53.8% | 56.8% | +3.0% |
| SOL | 52.3% | 55.1% | +2.8% |
| HIVE | 55.1% | 55.9% | +0.8% |
| AUCTION | 52.4% | 61.2% | +8.8% |
BTC and AUCTION showed the largest improvements — likely because they're more connected to broader market sentiment. Smaller altcoins like HIVE, which can move on coin-specific news, showed less improvement from macro data.
Timeframe Matters
We later tested macro features across 1-hour, 4-hour, and daily timeframes (Test K9):
| Timeframe | Improvement from Macro |
|---|---|
| 1h | +1.4% |
| 4h | +0.5% |
| 1d | +4.5% |
Daily timeframe showed the strongest macro effect. This makes intuitive sense — macro data is released daily, so daily candles naturally align with the data frequency. Hourly candles get the same macro value repeated 24 times (forward-filled), which dilutes the signal.
The FOMC Effect
One of our more interesting findings: the "FOMC week" binary feature (1 if a Federal Reserve meeting is within 3 days) was consistently in the top 20 most important features.
Markets behave differently around central bank meetings:
- Before: Volatility compresses as traders wait for the announcement
- During: Sharp directional moves as the decision is revealed
- After: Trend continuation or reversal depending on expectations vs reality
The model learned to adjust its predictions based on this calendar effect.
The Forward-Fill Problem
One caveat: macro data is daily, but we trade on hourly candles. We forward-fill daily values to match the hourly timestamps, which means the same macro value appears 24 times in a row.
This creates a subtle issue for some techniques. Our drift detection experiment (which checks if market conditions have changed) flagged the forward-filled macro data as "suspicious" — the repeating pattern is trivially detectable as a distributional artifact.
For the XGBoost model itself, this isn't a problem — the model handles repeated values fine. But it's worth noting for anyone replicating this approach.
Key Takeaways
- Macro data improves crypto predictions by 3-5%. This is one of the largest single improvements in our experiment series.
- Dollar strength is the strongest single predictor. DXY rate of change consistently ranks in the top features.
- Daily timeframe benefits most from macro features due to natural data alignment.
- BTC is most macro-sensitive. Makes sense — it's the most institutional, most correlated with traditional markets.
- The data is free. Economic data APIs are publicly available. No premium subscriptions required.
Part of Our Research Series
- 13,500 Model Fits Later: What Actually Works — Overview
- Why We Chose XGBoost Over LSTM — Model comparison
- This post — Macro features
- Meta-Labeling: Filtering Bad Trades — Signal quality
Full methodology: How Our AI Works
AI trading signals are probabilistic predictions, not financial advice. Macro correlations can break down without warning. Past performance does not guarantee future results.
Originally published at Nydar. Nydar is a free trading platform with AI-powered signals and analysis.
Top comments (0)