For intraday FX scalpers and execution desks, volatility forecasting is not academic: it determines quoting widths, position sizing, stop placement and whether a trade should be taken at all. Two broad families of approaches dominate practice today. On one side are parametric conditional heteroskedasticity models (GARCH and variants). On the other are non‑parametric realized‑volatility frameworks augmented with microstructure features drawn from limit order books and trade prints. This analysis compares the two in practical, implementation‑focused terms—what they do best, where they fail, and how to integrate forecasts into live execution in 2026 market structure.

Why intraday volatility forecasting matters for scalpers

Scalpers operate at horizons of seconds to minutes. A correct volatility forecast reduces three execution risks:

  • Adverse selection: tighter quotes during quiet regimes reduce earnings, while wider quotes during volatile windows protect against informed flow.
  • Slippage estimation: expected move size conditions position sizing and stop levels; underestimating volatility increases realized slippage and blowups.
  • Inventory and hedging: volatility dictates how quickly inventory should be neutralized and whether to use hedged options overlay for tail risk.

Consequently, forecast horizon, update frequency and latency are as important as raw accuracy.

GARCH family: strengths, weaknesses, and practical tips

GARCH models (GARCH(1,1), EGARCH, GJR‑GARCH) model conditional variance using a parsimonious parametric recursion. Strengths for FX scalpers:

  • Low data requirements: operate on low‑frequency returns (e.g., 1‑minute bars) and are robust to moderate data quality issues.
  • Clear statistical framework: parameters have interpretable persistence, leverage effects and long‑run variance.
  • Fast inference for low‑dimensional setups—suitable where computational resources are constrained.

Limitations:

  • Slow to adapt at the very short horizon. Parameter estimates smooth over fast regime shifts; shocks are incorporated only via the recursion.
  • Misspecified innovation distributions (fat tails, clustering of jumps) degrade probabilistic forecasts.
  • Neglect of order‑book dynamics: GARCH uses price returns only and ignores trade counts, imbalance and depth.

Practical tips if you choose GARCH:

  • Use high‑frequency returns (30s–1min) for intraday scalping, but beware microstructure noise—consider pre‑filtering returns with a median filter or subsampling.
  • Estimate parameters on rolling windows with short lookbacks (e.g., 1–5 trading days) to improve responsiveness.
  • Combine with heavy‑tailed innovation distributions (Student‑t, skew‑t) to get realistic VaR and extreme‑move estimates.

Realized‑volatility + microstructure: architecture and advantages

Realized‑volatility approaches estimate variance from intraday high‑frequency returns (sum of squared returns or robust bipower measures). Augmenting these with microstructure features—trade sign imbalance, trade count, quoted depth and spread—creates a model tailored for the seconds‑to‑minute horizon.

Why this family often outperforms GARCH for scalping:

  • Responsiveness: realized measures use the most recent ticks and microstructure signals, capturing volatility jumps almost immediately.
  • Feature richness: order‑book features provide early warning of liquidity dry‑ups or informed buying/selling that precede elevated volatility.
  • Flexible modeling: non‑parametric or machine‑learning regressors (random forests, gradient boosting, light neural nets) can map complex interactions between features and short‑term variance.

Key practical considerations:

  • Data requirements are high. You need clean tick data with trade sizes, trade direction (or signed ticks via tick rule or exchange flags), top‑of‑book depth and spreads.
  • Microstructure noise: naive realized estimators are biased by bid‑ask bounce and asynchronous timestamps; use subsampling or robust measures (e.g., realized kernel, bipower variation).
  • Latency matters: feature computation must be near real‑time. The architecture typically places feature engineers and model inference close to the matching engine or colocation providers to minimize stale signals.

Head‑to‑head: evaluation dimensions that matter

When choosing between a GARCH or realized‑microstructure approach, evaluate across these dimensions:

  1. Forecast horizon — GARCHs are serviceable at multi‑minute horizons; realized‑microstructure wins at seconds–1 minute.
  2. Update frequency and latency — Realized methods require continuous recomputation; GARCH can be updated less frequently.
  3. Robustness to data gaps — GARCH is more forgiving of occasional missing ticks or sparse trading; microstructure models degrade without full feed continuity.
  4. Interpretability — GARCH parameters are interpretable; black‑box ML on microstructure less so, complicating risk controls and regulatory explanations.
  5. Extreme event handling — Both struggle with unprecedented regime shifts; realized approaches can detect jumps faster, but may underpredict persistence of elevated volatility.

Hybrid and ensemble approaches: practical middle ground

Most desks no longer pick a single paradigm. Practical, deployable patterns in 2026 include:

  • Two‑tier forecasting: a realized‑microstructure model for immediate quoting decisions (sub‑minute) with a GARCH model providing a stabilizing long‑run variance estimate for inventory limits.
  • Meta‑ensembles: convex combination of forecasts where weights are time‑varying and driven by signal reliability metrics (e.g., feature completeness or recent forecast error).
  • Regime gating: use a quick regime classifier (based on volatility jumps and trade‑imbalance thresholds) to switch between conservative and aggressive forecast models.

Integrating forecasts into execution: concrete examples

Scalpers should translate variance forecasts into actionable controls:

  • Quoting width: map expected short‑term volatility to a quoting spread schedule with hysteresis to avoid quote flicker.
  • Order size scaling: scale aggressive order size as inverse root of forecast variance; cap maximum exposure during forecast spikes.
  • Stop and limit placement: adapt stop distances to predicted realized volatility rather than fixed pip amounts.
  • Inventory rebalancing cadence: increase hedge frequency when forecast variance exceeds a threshold to limit overnight gamma/delta leakage.

Operational constraints—latency, co‑location, throttling of market data—often force tradeoffs between the theoretical best predictor and the one practically deliverable. Realized‑microstructure models typically require investment in low‑latency data pipelines, whereas a tuned GARCH can run on a modest server with less operational overhead.

Practical recipe for scalpers and small execution desks

Follow a staged approach to adopt a volatility forecasting stack:

  1. Start with robust baseline: implement a short‑lookback GARCH(1,1) with Student‑t innovations on 1‑minute returns to set a conservative baseline.
  2. Collect and clean tick data: capture trades, quotes, sizes and NBBO where available; build microsecond‑aligned timestamps and handle outliers.
  3. Compute robust realized estimators: use bipower variation or realized kernels to estimate intraday variance and jump measures.
  4. Engineer microstructure features: recent trade count, signed volume imbalance, top‑of‑book depth ratio, spread percentile versus historical intraday profile.
  5. Prototype lightweight ML models: try gradient boosting with cross‑validation and walk‑forward testing focusing on mean squared error, directional accuracy for high‑volatility flags, and economic metrics (reduced slippage, improved P&L per trade).
  6. Backtest in a latency‑aware simulator: model market impact, queue fading and execution priority to assess real slippage—not just forecasting error.
  7. Progress to live A/B testing: run new model in passive quoting on a fraction of flow and compare realized slippage and P&L per trade.

Conclusion — tradeoffs and recommended approach

No single method dominates every use case. For pure ultra‑short‑horizon scalping, realized‑volatility augmented with microstructure signals usually provides faster and more informative forecasts—at the cost of higher data and ops complexity. GARCH‑type models remain valuable as a parsimonious, interpretable baseline and as a stabilizer for longer intraday horizons and risk limits. Best practice in 2026 is hybridization: let realized methods drive immediate quoting under normal data availability, but retain parametric models for fallback, regulatory transparency and long‑horizon risk metrics. Finally, evaluate success not only on forecast error but on economic outcomes—reduced slippage, improved hit rate and more stable inventory paths—because improved forecasts only matter if they improve the bottom line.