Update README.md
Browse files
README.md
CHANGED
|
@@ -23,23 +23,25 @@ Our model (with leakage) is deployed on the [TS-Arena benchmark](https://ts-aren
|
|
| 23 |
|
| 24 |
**Our model (without leakage) is competitive with the state-of-the-art notably compared to [MOIRAI2.0](https://arxiv.org/pdf/2511.11698)**
|
| 25 |
|
| 26 |
-
| Model
|
| 27 |
-
|---------
|
| 28 |
-
|
|
| 29 |
-
|
|
| 30 |
-
|
|
| 31 |
-
|
|
| 32 |
-
|
|
| 33 |
-
|
|
| 34 |
-
|
|
| 35 |
-
|
|
| 36 |
-
|
|
| 37 |
-
|
|
| 38 |
-
|
|
| 39 |
-
|
|
| 40 |
-
|
|
| 41 |
-
|
|
| 42 |
-
|
|
|
|
|
|
|
| 43 |
|
| 44 |
## Highlights
|
| 45 |
- Next-patch prediction objective (autoregressive, causal)
|
|
@@ -161,17 +163,6 @@ A common workaround is to feed only the median prediction (the 0.5 quantile) bac
|
|
| 161 |
|
| 162 |
An alternative approach is **autoregressive multi-quantile decoding**, as proposed in [Moirai 2.0](https://arxiv.org/pdf/2511.11698v1). This method enables consistent autoregressive generation while preserving the full predictive distribution across quantiles. However, it is computationally more expensive than the median-only approach as it requires duplicating the context for each quantile.
|
| 163 |
|
| 164 |
-
<div style="display: flex; gap: 10px; align-items: flex-start;">
|
| 165 |
-
<div>
|
| 166 |
-
<img src="assets/classic_forecast_animation.gif" alt="Classic Autoregressive Inference" width="450">
|
| 167 |
-
<p style="text-align:center;">Classic Autoregressive Inference</p>
|
| 168 |
-
</div>
|
| 169 |
-
<div>
|
| 170 |
-
<img src="assets/quantile_forecast_animation.gif" alt="Autoregressive Multi-Quantile Decoding" width="450">
|
| 171 |
-
<p style="text-align:center;">Autoregressive Multi-Quantile Decoding</p>
|
| 172 |
-
</div>
|
| 173 |
-
</div>
|
| 174 |
-
|
| 175 |
The algorithm proceeds as follows:
|
| 176 |
|
| 177 |
1. **Initialization**
|
|
|
|
| 23 |
|
| 24 |
**Our model (without leakage) is competitive with the state-of-the-art notably compared to [MOIRAI2.0](https://arxiv.org/pdf/2511.11698)**
|
| 25 |
|
| 26 |
+
| Model | Win Rate | Skill Score |
|
| 27 |
+
|---------|----------:|------------:|
|
| 28 |
+
| Chronos-2 | 0.89 | 0.47 |
|
| 29 |
+
| TimesFM-2.5 | 0.81 | 0.47 |
|
| 30 |
+
| TiRex | 0.81 | 0.43 |
|
| 31 |
+
| FlowState | 0.71 | 0.42 |
|
| 32 |
+
| Toto-1.0 | 0.70 | 0.41 |
|
| 33 |
+
| **PatchFM** | 0.69 | 0.40 |
|
| 34 |
+
| Moirai-2.0 | 0.65 | 0.40 |
|
| 35 |
+
| Chronos-Bolt | 0.62 | 0.39 |
|
| 36 |
+
| TFT | 0.47 | 0.32 |
|
| 37 |
+
| Sundial-Base | 0.43 | 0.34 |
|
| 38 |
+
| PatchTST | 0.42 | 0.30 |
|
| 39 |
+
| CatBoost | 0.32 | 0.23 |
|
| 40 |
+
| LightGBM | 0.29 | 0.21 |
|
| 41 |
+
| AutoTheta | 0.28 | 0.05 |
|
| 42 |
+
| Seasonal Naive | 0.19 | 0.00 |
|
| 43 |
+
| Naive | 0.13 | -0.45 |
|
| 44 |
+
| Drift | 0.09 | -0.46 |
|
| 45 |
|
| 46 |
## Highlights
|
| 47 |
- Next-patch prediction objective (autoregressive, causal)
|
|
|
|
| 163 |
|
| 164 |
An alternative approach is **autoregressive multi-quantile decoding**, as proposed in [Moirai 2.0](https://arxiv.org/pdf/2511.11698v1). This method enables consistent autoregressive generation while preserving the full predictive distribution across quantiles. However, it is computationally more expensive than the median-only approach as it requires duplicating the context for each quantile.
|
| 165 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 166 |
The algorithm proceeds as follows:
|
| 167 |
|
| 168 |
1. **Initialization**
|