Novel Time Series Features for Intraday Financial Forecasting

Source: Notion | Last edited: 2025-03-19 | ID: 1b02d2dc-3ef...

---

• The Bollinger Bandwidth Delta (BBD)

Assuming no constraints on computational efficiency or interpretability, propose low-hanging but novel, high-impact, lesser-known time series features for financial forecasting for intraday crypto, futures, and/or forex (but not options).

Only interested in features derived from price and volume, but not order book data and alternative sources (e.g., sentiment, news, macroeconomic indicators).

GPT-4.5 Deep Research

Here are low-hanging but novel high-impact less-known time series features for intraday crypto, futures, and forex forecasting, focusing only on price and volume data:

---

1. Volume-Weighted Intraday Trend Strength (VWITS)

• Measures how much price moves in the dominant direction weighted by volume.
• Helps identify strong trends where price moves are backed by high volume.
• Formula:

where is price change, is volume, and is the period.

$$VWITS=∑t=1Tsign(ΔPt)×VtVWITS = \sum_{t=1}^{T} \text{sign}(\Delta P_t) \times V_t$$ $$ΔPt\Delta P_t$$ $$VtV_t$$ $$TT$$

---

2. Volatility-of-Volatility (VoV)

• Measures how volatile the volatility itself is.
• Useful for detecting upcoming price bursts.
• Formula:

$$VoV=std(log-return volatility over a rolling window)VoV = \text{std}(\text{log-return volatility over a rolling window})$$

---

3. Volume Acceleration

• Measures the rate of change of volume to detect impending large moves.
• Formula:

$$Vaccel=Vt−Vt−1Vt−1V_{\text{accel}} = \frac{V_t - V_{t-1}}{V_{t-1}}$$

• High acceleration precedes breakouts.

---

4. Price Compression Ratio (PCR)

• Identifies tight consolidation ranges that often lead to breakouts.
• Formula:

$$PCR=max⁡(Pt)−min⁡(Pt)ATRPCR = \frac{\max(P_t) - \min(P_t)}{\text{ATR}}$$

• A low value suggests price is compressing before an expansion.

---

5. Intraday Volume Shock Index (IVSI)

• Detects sudden spikes in volume compared to recent norms.
• Formula:

where and are the mean and std of volume over a rolling window.

$$IVSI=Vt−μVσVIVSI = \frac{V_t - \mu_{V}}{\sigma_{V}}$$

μV\mu_V

σV\sigma_V

• High IVSI suggests large player activity before price moves.

---

6. Inverse Trend Momentum (ITM)

• Captures how much price moves against the dominant trend.
• Helps detect weakening trends before reversals.
• Formula:

$$ITM=∑t=1T−sign(ΔPt)×∣Pt−Pt−1∣ITM = \sum_{t=1}^{T} -\text{sign}(\Delta P_t) \times \left| P_t - P_{t-1} \right|$$

• A rising ITM suggests trend exhaustion.

---

7. Fractional Differentiation of Price (FDP)

• Captures long-memory effects in price movements.
• Formula:

where are coefficients derived from fractional differencing.

$$Ptd=∑k=0∞ckPt−kP^d_t = \sum_{k=0}^{\infty} c_k P_{t-k}$$ $$ckc_k$$

• Helps retain memory while making the series stationary.

---

8. Volume-Weighted Price Reversion (VWPR)

• Measures how much price reverts relative to the previous volume profile.
• Formula:

$$VWPR=Pt−PVWAPPVWAPVWPR = \frac{P_t - P_{VWAP}}{P_{VWAP}}$$

• Helps in mean-reversion strategies.

---

9. High-Low Asymmetry Index (HLAI)

• Quantifies whether intraday highs or lows are more dominant.
• Formula:

$$HLAI=∑(Ht−Ct−1)−∑(Ct−1−Lt)∑(Ht−Lt)HLAI = \frac{\sum (H_t - C_{t-1}) - \sum (C_{t-1} - L_t)}{\sum (H_t - L_t)}$$

• A high value suggests bullish dominance, and low value suggests bearish pressure.

---

10. Volume Density Clustering (VDC)

• Identifies volume accumulation zones where price sticks.
• Computed as time spent within a percentage range of the price weighted by volume.
• Helps identify key support and resistance areas.

---

11. Price Gradient Curvature (PGC)

• Measures the curvature of the price trend.
• Formula:

$$PGC=d2Ptdt2PGC = \frac{d^2P_t}{dt^2}$$

• High curvature values indicate sharp trend shifts.

---

12. Bollinger Band Slope Momentum (BBSM)

• Measures how fast Bollinger Bands are widening/narrowing.
• Formula:

$$BBSM=σt−σt−1σt−1BBSM = \frac{\sigma_t - \sigma_{t-1}}{\sigma_{t-1}}$$

• Helps detect volatility expansion before breakouts.

---

13. Price Distribution Skewness (PDS)

• Identifies whether price movements are more upward or downward skewed.
• Formula:

$$PDS=E[(Pt−μP)3]σP3PDS = \frac{E[(P_t - \mu_P)^3]}{\sigma_P^3}$$

• Helps detect bias towards bullish or bearish price action.

---

14. Range Expansion Score (RES)

• Quantifies how much range expansion is happening compared to normal.
• Formula:

$$RES=Current range (H-L)Rolling average rangeRES = \frac{\text{Current range (H-L)}}{\text{Rolling average range}}$$

• High values indicate volatility breakout potential.

---

15. Delta-Volume Weighted Average Price (Delta-VWAP)

• Captures how much price moves relative to volume changes.
• Formula:

$$Delta-VWAP=∑t=1T(Pt−PVWAP)×Vt\text{Delta-VWAP} = \sum_{t=1}^{T} (P_t - P_{VWAP}) \times V_t$$

• Helps in detecting volume-driven price imbalances.

---

Conclusion

These features combine price and volume to extract novel insights for intraday trading in crypto, forex, and futures. Several features are momentum-based, mean-reversion-based, or volatility-expansion-based, providing low-hanging but impactful signals.

Which features would you like more details on, including implementation examples?

DeepSeek R1

Here’s a merged, prioritized list of novel, high-impact time series features combining ideas from both previous responses, tailored for intraday crypto/futures/forex forecasting using price and volume data only. These features emphasize regime shifts, nonlinear dynamics, multiscale patterns, and liquidity-volume interactions, while avoiding redundancy:

---

I. Multiscale Complexity & Predictability

1. Wavelet Multiscale Sample Entropy

• Compute sample entropy across wavelet-decomposed scales (e.g., 5-min vs. 1-hour).
• Use: Detects shifts in market efficiency at different frequencies.

1. Multifractal Spectrum Width (MF-DFA)

• Quantify heterogeneity in scaling behavior via multifractal detrended fluctuation analysis.
• Use: Wider spectrum → chaotic regimes (e.g., crypto flash crashes).

1. VMD Energy Ratios

• Energy distribution of variational mode decomposition (VMD) components.
• Use: Dominant high-frequency energy → noise; low-frequency → trend persistence.

1. Hilbert Instantaneous Phase

• Phase angles of price cycles via the Hilbert transform.
• Use: Phase inversion → trend reversal signals (e.g., forex intraday cycles).

---

II. Nonlinear Dependencies & Asymmetries

1. Mutual Information Lags

• Nonlinear predictive power between lagged returns and future price, estimated via k-NN.
• Use: Captures hidden lead-lag relationships (e.g., BTC futures → altcoin rallies).

1. Quantile Transition Probabilities

• Likelihood of returns jumping between extreme quantiles (e.g., 5th → 95th percentile).
• Use: Flags tail-risk regimes (e.g., “black swan” precursors in futures).

1. Autocorrelation Asymmetry

• Difference in autocorrelation of positive vs. negative returns.
• Use: Momentum/reversal bias (e.g., ρ₊ > ρ₋ → trending markets).

1. Entropic Causality (Transfer Entropy)

• Information flow between volume and price series.
• Use: High TE(volume→price) → impending volatility (e.g., crypto pumps/dumps).

---

III. Regime Shift Detection

1. Hurst Exponent + Variance Ratio Fusion

• Combine Hurst (R/S) and variance ratio (VR) to detect mean-reversion/momentum conflicts.
• Use: VR < 1 but H > 0.5 → conflicting signals → regime transition.

1. Recurrence Plot Metrics (RQA)

• Determinism, laminarity, and entropy from recurrence quantification analysis.
• Use: High determinism → predictable patterns; low laminarity → instability.

1. Quantile Autoregression Residuals

• Residuals from extreme quantile (e.g., 90th) autoregressive models.
• Use: Large residuals → overreaction reversals (e.g., forex news spikes).

---

IV. Volume-Price Integration

1. Volume-Weighted Fractal Dimension

• Higuchi’s fractal dimension scaled by normalized volume.
• Use: High fractal dim + high volume → chaotic liquidity-driven moves.

1. Amihud Illiquidity Acceleration

• Rate of change of Amihud’s metric: $\frac{d}{dt} \left( \frac{|r_t|}{\text{Volume}_t} \right)$.
• Use: Spiking acceleration → liquidity crises (e.g., crypto illiquid pairs).

1. Spectral Risk-Volume Coupling

• Coherence between volume changes and Fourier power of returns.
• Use: High coherence at low frequencies → persistent volume-driven risks.

---

V. Pattern Recognition & Latent Structure

1. Temporal Convolution Kernels

• Activations from shallow CNNs trained on raw returns/volume sequences.
• Use: Automatically detects recurring intraday motifs (e.g., futures rollover patterns).

1. Neural ODE Latent States

• Hidden states from neural ordinary differential equations modeling price dynamics.
• Use: Encodes nonstationary drift/diffusion terms (e.g., crypto hyper-volatility).

1. Sparse Symbolic Representations

• Encode price/volume into symbolic sequences (SAX) + rare pattern counts.
• Use: Flags anomalous intraday sequences (e.g., “stealth” accumulation phases).

---

VI. Implementation Tips

• Fusion: Combine entropy/volatility metrics (I) with regime detectors (III) for robustness.
• Nonlinearity First: Prioritize mutual information (6) over linear autocorrelation.
• Volume Context: Always pair price features (e.g., fractal dim) with volume-weighted versions.
• Open-Source Tools:
  • pywt (wavelets), nolds (entropy), MFDFA (multifractals), tsfresh (feature engineering).

---

Why These Work

1. Antifragility: Metrics like quantile transitions (6) and RQA (10) thrive in volatile regimes.
2. Adaptivity: Wavelet/VMD (1,3) adjust to intraday seasonality without fixed windows.
3. Volume as Signal: Most models treat volume as noise; here, it’s a core driver (12,13).
4. Low Correlation: Features span entropy (info theory), geometry (fractals), and dynamics (ODEs). This toolkit avoids overused features (RSI, MACD) and focuses on underappreciated drivers of intraday chaos, such as liquidity-driven fractal roughness or phase-space determinism.