Seminarium z udziałem Jana Wołyniaka, studenta WNE UW
Zapraszamy na 71. seminarium, organizowane wspólnie przez ośrodki badawcze QFRG oraz DSLab na Wydziale Nauk Ekonomicznych Uniwersytetu Warszawskiego.
Podczas spotkania Jan Wołyniak (Quantitative Finance Research Group) omówi wyniki badania pt. „Zero-Shot Transformer Models for High-Frequency Financial Time Series Forecasting in Trading Applications”. Jego współautorem jest dr hab. Robert Ślepaczuk, prof. ucz. (Quantitative Finance Research Group, Katedra Finansów Ilościowych i Uczenia Maszynowego WNE).
Autorzy badania oceniają zdolność wielkich modeli transformatywnych (time-series foundation models) do prognozowania i generowania sygnałów inwestycyjnych oraz sprawdzają, jak radzą sobie one w zmiennym środowisku finansowym.
Na seminarium zapraszamy 8 czerwca o godz. 18:30. Spotkanie odbędzie się w formie hybrydowej: stacjonarnie w sali B002 (WNE UW) oraz zdalnie - na platformie Zoom. W celu otrzymania linku do spotkania prosimy o kontakt na adres &?k/42Zn9P-u.GRmvxc8aH]#[t.[w~pEPTHv[79Jv\]YER1.
Zainteresowanych wysłuchaniem wykładu prosimy o obecność możliwie na 5 minut przed jego rozpoczęciem. Prelekcja potrwa 45 min, a po niej - zwyczajowo, przewidziana jest dyskusja.
Zachęcamy do zapoznania się z abstraktem badania, dostępnym poniżej.
---------------------------
Abstrakt:
Recent advances in machine learning have led to the development of time-series foundation models - large transformer architectures trained on massive collections of heterogeneous temporal data. Models such as Chronos, TimesFM, Lag-Llama, and Moirai aim to learn general temporal structures and perform forecasting in a zero-shot setting without task-specific retraining. These models promise broad applicability across domains, including demand forecasting, energy systems, and financial markets. Financial time series present a particularly challenging forecasting environment. Asset prices exhibit high stochasticity, heavy-tailed distributions, and extremely low signal-to-noise ratios. According to the Efficient Market Hypothesis, short-term price movements should be difficult to predict because the available information is rapidly incorporated into market prices. Despite this, machine learning methods, including recurrent neural networks such as LSTM, have been widely explored in financial forecasting with mixed results. This study examines whether zero-shot transformer forecasting models can generate trading signals from high-frequency financial data. The analysis uses hourly Bitcoin prices from 2018-01-01 through 2021-01-29, with auxiliary variables including gold, the S&P 500, VIX, and macroeconomic indicators, under both univariate and multivariate settings. We evaluate several open-source time series foundation models, including Chronos, Chronos-2, TimesFM, TOTO, Moirai, Lag-Llama, MOMENT, and Tiny Time Mixers. Model forecasts are incorporated into trading signals using a threshold-based decision rule that generates long/short/flat signals. Strategies are evaluated through a consistent backtesting pipeline using risk-adjusted performance metrics. The results show that transaction costs fundamentally alter the model's performance. In frictionless simulations, several models achieve extremely high cumulative returns due to very frequent trading. However, incorporating realistic transaction costs naturally leads strategies to focus on trades with higher-confidence predictions. In this setting, models producing selective signals outperform highly active strategies. Overall, the results suggest that while time-series foundation models can occasionally generate profitable signals, robust performance in high-frequency financial markets likely requires signal sparsity and domain-adapted or hybrid modeling approaches.