File size: 4,903 Bytes
3335f2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
---
license: apache-2.0
pretty_name: FUTURE-TS
tags:
- time-series
- forecasting
- benchmark
- evaluation
- foundation-models
- leaderboard
task_categories:
- time-series-forecasting
---

# FUTURE-TS

## Summary

FUTURE-TS v0.1.0 is a public-preview benchmark for time-series foundation
models. It treats evaluation as an executable protocol: task cards declare
issue times, horizons, available history, delayed labels, adaptation budgets,
resource limits, and metrics; submissions are validated, scored, and audited
against those contracts.

The core rule is future-only ranking: a submission is only ranked on labels
that were not available when the model was frozen.

## Benchmark Surface

- Release: `v0.1.0`
- Strict surface: `benchmarks/v1/benchmark.json`
- Surface identifier: `future-ts-v1`
- Tasks: 25
- Tiers: `public_dev`, `blind_archive`, `live`
- Tracks: forecasting core, covariate-aware, multivariate relational, data
  quality, event, transfer, multimodal context
- Required manifest: `require_pretraining_manifest=true`
- Headline score: `tier_weighted_score`
- Additional views: capability vector `(F, U, R, A, E, D)`, Pareto flag,
  rank-based mean-of-ranks aggregate with bootstrap CI

The strict surface rejects manifestless submissions so "clean" and
"undeclared" are not conflated.

## Current Empirical Run

The current canonical empirical artifacts are in:

```text
reports/tsfm_ai_empirical_v2_multi_budget/
```

This run is the `future-ts-empirical-v2` hosted TSFM.ai slice used by the
empirical paper. It exercises 15 real-data tasks across three context budgets:

| Budget | Context length |
|--------|----------------|
| `zero_shot` | 96 |
| `few_shot` | 192 |
| `s16` | 288 |

Current result snapshot:

- Surfaced public catalog entries: 52
- Merged public submissions: 51
- Scored public models: 47
- Positive overall scores: 30
- Current winner: `Datadog/Toto-2.0-1B`
- Winner score: `0.2369`
- Rank-consistency leader by mean rank: `amazon/chronos-2`

Top five by tier-weighted score:

| Rank | Model | Score |
|------|-------|-------|
| 1 | `Datadog/Toto-2.0-1B` | 0.2369 |
| 2 | `Datadog/Toto-2.0-313m` | 0.2214 |
| 3 | `NX-AI/TiRex` | 0.2024 |
| 4 | `Salesforce/moirai-2.0-R-small` | 0.1860 |
| 5 | `amazon/chronos-2` | 0.1851 |

## How To Submit A Model Entry

Submit a pull request under:

```text
submissions/community/<org>_<model>/
```

Each submission directory must contain:

- `script.py`: sealed-runner entry point
- `declaration.json`: metadata, declarations, artifact URI, and pretraining
  manifest
- `README.md`: short model description and links

CI validates the directory structure and smoke-runs `script.py` through the
sealed runner. After review, the evaluator runs the full benchmark and can
publish the resulting `BenchmarkReport`.

See [submission-guide.md](submission-guide.md) for the full process.

## What This Does Not Yet Claim

The v0.1.0 release is a local benchmark package plus sealed-runner MVP. It is
not yet a fully hosted, externally attested live benchmark service.

Do not read the current empirical run as a permanent universal TSFM ranking.
It is a first real-data slice over 15 tasks and one materialized wave. Wider
task coverage, repeated waves, hosted attestation, immutable submission
windows, and stronger manifest evidence are the next steps.

See [validity-envelope.md](validity-envelope.md) for the precise claim
boundary.

## Local Commands

Validate the strict benchmark:

```bash
python3 -m future_ts.cli validate-benchmark benchmarks/v1/benchmark.json
```

Run the sealed reference submission:

```bash
python3 -m future_ts.cli run-sealed \
  examples/submissions/reference_seasonal_naive.py \
  .tmp/task_windows.json \
  --output .tmp/submission.json \
  --submission-id local::reference_seasonal_naive \
  --model-name "Reference Seasonal Naive"
```

Build a leaderboard from saved reports:

```bash
python3 -m future_ts.cli leaderboard reports/tsfm_ai_empirical_v2_multi_budget/*.report.json
```

These core workflows do not require TSFM.ai. A TSFM.ai API key is only needed
to run the optional hosted-catalog evaluation commands that invoke hosted
models.

## More Detail

- Design paper: `paper/future_ts_design.pdf`
- Empirical paper: `paper/future_ts_empirical.pdf`
- Benchmark design notes: [benchmark-design.md](benchmark-design.md)
- Submission guide: [submission-guide.md](submission-guide.md)
- Sealed runner: [sealed-runner.md](sealed-runner.md)
- Validity envelope: [validity-envelope.md](validity-envelope.md)


## Repository Contents

- `benchmarks/`: public benchmark surfaces and task-card references.
- `schemas/`: JSON Schemas for benchmark, task, submission, and actual records.
- `docs/`: benchmark card, submission guide, sealed-runner notes, and validity envelope.
- `reports/tsfm_ai_empirical_v2_multi_budget/`: canonical public empirical reports.
- `paper/`: design and empirical PDFs.