Horizon-Specific Volatility Forecasting: Why Model Fit Depends on Conditions
AI-assisted research by PyInvesting. Sources and limitations are provided below. Educational content; not personalized investment advice.
AI-assisted article drafted and checked with AI against the sources below. Educational information, not personal investment advice.
How does forecast horizon affect model comparisons?
The Federal Reserve study reports different leading models across horizons: Markov-switching HAR performs best at short horizons, ARFIMA generally leads at the monthly horizon, and the five-day horizon is intermediate. [S7]
Do machine-learning models consistently lead?
In the Federal Reserve comparison, machine-learning models sometimes improved on HAR but did not systematically outperform the broader group of econometric models. [S7]
Why should evaluations account for changing regimes?
The arXiv market-making study describes stationary order-flow assumptions as vulnerable at realistic microstructure timescales and reports that persistent directional imbalance can produce large drawdowns for a stationarily trained strategy. [S4]
What this means
Questions for the reader: Does the model’s evaluation horizon match its intended use? Does the test examine persistence and nonlinear dynamics rather than only average performance? Does it include regime changes or persistent directional conditions? What failure mode could emerge if actual conditions depart from the model’s assumptions?
Limitations
The supplied excerpts cover different applications: the Federal Reserve source examines realized-volatility forecasting, while the arXiv source examines market making. They are used here to frame model-evaluation questions, not to claim that their results are directly comparable or that either study establishes a universal investment rule. The excerpts do not provide a complete portfolio implementation analysis.
evaluating an investment strategy
Sources
- Deep Learning of Robust Market Making under Regime-Switching Order Flow — 2026-09-11T04:00:00+00:00
- Linear and Nonlinear Econometric Models versus Machine-Learning Models: Evidence from Realized-Volatility Forecasting — 2026-09-02T15:30:00+00:00