Spaces:
Running
Running
Update README.md
Browse files
README.md
CHANGED
|
@@ -14,7 +14,7 @@ pinned: false
|
|
| 14 |
|
| 15 |
<div style="padding: 12px; border-radius: 10px; box-shadow: 0 4px 8px rgba(0, 0, 0, 0.1);">
|
| 16 |
|
| 17 |
-
## 🚀
|
| 18 |
|
| 19 |
We are pleased to introduce **FinText-TSFM**, a comprehensive suite of **time series foundation models (TSFMs)** developed for financial forecasting and quantitative research.
|
| 20 |
This release accompanies the paper
|
|
@@ -35,7 +35,6 @@ by *Eghbal Rahimikia, Hao Ni, and Weiguan Wang (2025)*.
|
|
| 35 |
This release includes variants of **Chronos** and **TimesFM** architectures adapted for financial time series:
|
| 36 |
- Chronos-Tiny / Mini / Small
|
| 37 |
- TimesFM-8M / 20M
|
| 38 |
-
- Parameter counts range from **8M to 200M+**.
|
| 39 |
|
| 40 |
- **Performance Insights:**
|
| 41 |
Our findings show that **off-the-shelf TSFMs** underperform in zero-shot forecasting, while **finance-pretrained models** achieve large gains in both predictive accuracy and portfolio Sharpe ratios.
|
|
@@ -43,8 +42,6 @@ by *Eghbal Rahimikia, Hao Ni, and Weiguan Wang (2025)*.
|
|
| 43 |
- **Evaluation Scope:**
|
| 44 |
Models are benchmarked across **U.S. and international equities**, using rolling windows (5, 21, 252, 512 days) and **18M+ out-of-sample forecasts**.
|
| 45 |
|
| 46 |
-
- **Open Science Commitment:**
|
| 47 |
-
All released models are available in **FP32** format for full transparency and reproducibility.
|
| 48 |
|
| 49 |
---
|
| 50 |
|
|
@@ -52,9 +49,7 @@ by *Eghbal Rahimikia, Hao Ni, and Weiguan Wang (2025)*.
|
|
| 52 |
|
| 53 |
- **Architecture:** Transformer-based TSFMs (Chronos & TimesFM)
|
| 54 |
- **Training Regime:** Pre-training from scratch, fine-tuning, and zero-shot evaluation
|
| 55 |
-
- **Objective:** Mean squared error (MSE) for continuous returns; cross-entropy for tokenized sequences
|
| 56 |
- **Compute:** >50,000 GPU hours on NVIDIA GH200 Grace Hopper clusters
|
| 57 |
-
- **Data Sources:** CRSP, Compustat Global, JKP factors, and proprietary merged panels (1990–2023)
|
| 58 |
|
| 59 |
---
|
| 60 |
|
|
@@ -74,11 +69,12 @@ Please cite the accompanying paper if you use these models:
|
|
| 74 |
|
| 75 |
This project was made possible through computational and institutional support from:
|
| 76 |
- **Isambard-AI National AI Research Resource (AIRR)**
|
| 77 |
-
- **The University of Manchester** (Research IT & Computational Shared Facility)
|
|
|
|
| 78 |
- **N8 Centre of Excellence in Computationally Intensive Research (N8 CIR)** — EPSRC Grant EP/T022167/1
|
| 79 |
- **University College London** and **Shanghai University**
|
| 80 |
- **The Alan Turing Institute**
|
| 81 |
-
|
| 82 |
|
| 83 |
---
|
| 84 |
|
|
|
|
| 14 |
|
| 15 |
<div style="padding: 12px; border-radius: 10px; box-shadow: 0 4px 8px rgba(0, 0, 0, 0.1);">
|
| 16 |
|
| 17 |
+
## 🚀 TSFMs Release
|
| 18 |
|
| 19 |
We are pleased to introduce **FinText-TSFM**, a comprehensive suite of **time series foundation models (TSFMs)** developed for financial forecasting and quantitative research.
|
| 20 |
This release accompanies the paper
|
|
|
|
| 35 |
This release includes variants of **Chronos** and **TimesFM** architectures adapted for financial time series:
|
| 36 |
- Chronos-Tiny / Mini / Small
|
| 37 |
- TimesFM-8M / 20M
|
|
|
|
| 38 |
|
| 39 |
- **Performance Insights:**
|
| 40 |
Our findings show that **off-the-shelf TSFMs** underperform in zero-shot forecasting, while **finance-pretrained models** achieve large gains in both predictive accuracy and portfolio Sharpe ratios.
|
|
|
|
| 42 |
- **Evaluation Scope:**
|
| 43 |
Models are benchmarked across **U.S. and international equities**, using rolling windows (5, 21, 252, 512 days) and **18M+ out-of-sample forecasts**.
|
| 44 |
|
|
|
|
|
|
|
| 45 |
|
| 46 |
---
|
| 47 |
|
|
|
|
| 49 |
|
| 50 |
- **Architecture:** Transformer-based TSFMs (Chronos & TimesFM)
|
| 51 |
- **Training Regime:** Pre-training from scratch, fine-tuning, and zero-shot evaluation
|
|
|
|
| 52 |
- **Compute:** >50,000 GPU hours on NVIDIA GH200 Grace Hopper clusters
|
|
|
|
| 53 |
|
| 54 |
---
|
| 55 |
|
|
|
|
| 69 |
|
| 70 |
This project was made possible through computational and institutional support from:
|
| 71 |
- **Isambard-AI National AI Research Resource (AIRR)**
|
| 72 |
+
- **The University of Manchester** (Research IT & Computational Shared Facility)
|
| 73 |
+
- **Alliance Manchester Business School (AMBS), University of Manchester**
|
| 74 |
- **N8 Centre of Excellence in Computationally Intensive Research (N8 CIR)** — EPSRC Grant EP/T022167/1
|
| 75 |
- **University College London** and **Shanghai University**
|
| 76 |
- **The Alan Turing Institute**
|
| 77 |
+
|
| 78 |
|
| 79 |
---
|
| 80 |
|