Regime-Based Nasdaq Futures Trading: LSTM vs Transformer vs Buy-and-Hold
Author
Mähleke, Niklas
Term
4. semester
Publication year
2025
Submitted on
2025-06-02
Pages
62
Abstract
Kortsigtede tradere og risikomanagere mangler ofte pålidelige værktøjer fra minut til minut, fordi traditionelle tidsseriemodeller kan bryde sammen, når likviditeten (hvor let det er at handle) og volatiliteten (hvor meget priserne svinger) ændrer sig hurtigt. Denne afhandling afprøver en enkel regel: Der handles kun, når et klart markedsregime er identificeret—"bull" (optrend), "bear" (nedtrend) eller "sidelæns". Jeg klassificerede NASDAQ-futures i ét-minuts intervaller fra januar 2015 til juni 2024 i disse regimer og trænede to deep-learning-modeller til at forudsige dem. En LSTM (Long Short-Term Memory) er et neuralt netværk, der lærer, hvad der skal huskes og glemmes over tid, nyttigt når både helt nye og lidt ældre prismønstre betyder noget. En Transformer er en anden netværksarkitektur, der bruger en attention-mekanisme til at fokusere på de mest relevante tidligere datapunkter for hver ny forudsigelse. I out-of-sample tests på data fra juli–december 2024 identificerede begge modeller reelle regimeskift inden for cirka ti minutter i gennemsnit og overgik en simpel køb-og-behold-strategi. De højere afkast kom dog med større udsving, så signalerne indebærer mere risiko. Samlet set viser en ramme på minuttiveau med deep learning, at LSTM- og Transformer-signaler kan slå køb-og-behold, hvis man accepterer større op- og nedture.
Short-term traders and risk managers often lack reliable minute-by-minute tools because traditional time-series models can fail when liquidity (how easy it is to trade) and volatility (how much prices move) change quickly. This thesis tests a simple rule: only trade when a clear market regime is detected—"bull" (uptrend), "bear" (downtrend), or "sideways". I labeled one-minute NASDAQ futures bars from January 2015 to June 2024 with these regimes and trained two deep-learning models to predict them. An LSTM (Long Short-Term Memory) is a neural network that learns what to remember or forget over time, useful when both recent and slightly older price patterns matter. A Transformer is another neural network design that uses an attention mechanism to focus on the most relevant past data for each new prediction. In out-of-sample tests on July–December 2024 data, both models identified real regime shifts within about ten minutes on average and outperformed a simple buy-and-hold strategy. However, the higher returns came with larger swings, so these signals involve more risk. Overall, this minute-level deep-learning framework shows that LSTM and Transformer signals can beat buy-and-hold if investors are willing to accept bigger ups and downs.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Documents