File size: 10,253 Bytes
2053767
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ea1a58e
 
 
9fba3c2
 
 
 
 
2053767
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ea1a58e
 
 
 
 
 
2053767
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
85698b8
 
 
 
2053767
 
 
 
 
 
 
 
 
9fba3c2
 
 
 
2053767
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9fba3c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2053767
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9fba3c2
 
 
 
 
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
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
from __future__ import annotations

import math
import os
from statistics import mean
from typing import Any

from schemas import HealthResponse, PredictRequest, PredictResponse, PredictionItem


class TimesFmService:
    """HF Space service wrapper for TimesFM.



    By default this service attempts real HuggingFace CPU inference. If model

    loading fails and `TIMESFM_ALLOW_BASELINE_FALLBACK=true`, it falls back to

    the deterministic baseline implementation.

    """

    def __init__(self) -> None:
        self.model_id = "timesfm"
        self.model_name = os.getenv(
            "TIMESFM_MODEL_NAME",
            "google/timesfm-2.5-200m-transformers",
        )
        self.backend = os.getenv("TIMESFM_BACKEND", "hf_cpu").strip() or "hf_cpu"
        self.device = "cpu"
        self.runtime_revision = os.getenv("TIMESFM_RUNTIME_REVISION", "timesfm-hf-patch-align-v1")
        self.max_context_length = int(os.getenv("TIMESFM_MAX_CONTEXT_LENGTH", "512"))
        self.max_horizon_step = int(os.getenv("TIMESFM_MAX_HORIZON_STEP", "288"))
        self.patch_length = int(os.getenv("TIMESFM_PATCH_LENGTH", "32"))
        self.confidence_floor = float(os.getenv("TIMESFM_CONFIDENCE_FLOOR", "0.20"))
        self.confidence_ceiling = float(os.getenv("TIMESFM_CONFIDENCE_CEILING", "0.85"))
        self.min_required_points = int(os.getenv("TIMESFM_MIN_REQUIRED_POINTS", "32"))
        self.allow_baseline_fallback = os.getenv("TIMESFM_ALLOW_BASELINE_FALLBACK", "false").lower() == "true"

        self.ready = False
        self.load_error = ""
        self._torch = None
        self._model = None

        self._initialize_backend()

    def health(self) -> HealthResponse:
        return HealthResponse(
            status="ok" if self.ready else "degraded",
            model=self.model_name,
            model_id=self.model_id,
            backend=self.backend,
            device=self.device,
            ready=self.ready,
            max_context_length=self.max_context_length,
            max_horizon_step=self.max_horizon_step,
            patch_length=self.patch_length,
            runtime_revision=self.runtime_revision,
        )

    def predict(self, payload: PredictRequest) -> PredictResponse:
        self._validate_request(payload)
        closes = payload.close_prices[-payload.context_length :]

        if self.backend == "hf_cpu":
            if not self.ready:
                raise RuntimeError(self.load_error or "timesfm backend not ready")
            predictions = self._predict_with_hf(closes, payload.horizons)
        else:
            predictions = self._predict_with_baseline(closes, payload.horizons)

        return PredictResponse(model_id=self.model_id, predictions=predictions)

    def _initialize_backend(self) -> None:
        if self.backend == "baseline_cpu":
            self.ready = True
            return

        if self.backend != "hf_cpu":
            raise ValueError(f"unsupported TIMESFM_BACKEND={self.backend}")

        try:
            self._load_hf_model()
            self.ready = True
        except Exception as exc:
            self.load_error = f"timesfm hf load failed: {exc}"
            if self.allow_baseline_fallback:
                self.backend = "baseline_cpu"
                self.ready = True
            else:
                self.ready = False

    def _load_hf_model(self) -> None:
        import torch
        from transformers import TimesFm2_5ModelForPrediction

        self._torch = torch
        torch.set_num_threads(max(1, int(os.getenv("TIMESFM_TORCH_THREADS", "2"))))
        self._model = TimesFm2_5ModelForPrediction.from_pretrained(
            self.model_name,
            torch_dtype=torch.float32,
        )
        self._model.to("cpu")
        self._model.eval()

    def _predict_with_hf(

        self, close_prices: list[float], horizons: list[int]

    ) -> list[PredictionItem]:
        assert self._torch is not None
        assert self._model is not None

        torch = self._torch
        context = self._aligned_hf_context(close_prices)
        past_values = [torch.tensor(context, dtype=torch.float32)]
        freq = torch.tensor([0], dtype=torch.long)

        with torch.inference_mode():
            outputs = self._model(
                past_values=past_values,
                freq=freq,
                forecast_context_len=len(context),
                return_forecast_on_context=False,
            )

        dense_mean = outputs.mean_predictions[0].detach().cpu().tolist()
        if len(dense_mean) < max(horizons):
            raise RuntimeError(
                f"TimesFM output horizon {len(dense_mean)} is shorter than requested {max(horizons)}"
            )

        dense_conf = self._timesfm_confidence(outputs, dense_mean)
        predictions: list[PredictionItem] = []
        for step in horizons:
            predictions.append(
                PredictionItem(
                    step=step,
                    pred_price=round(max(0.00000001, float(dense_mean[step - 1])), 8),
                    pred_confidence=round(dense_conf[step - 1], 4),
                )
            )
        return predictions

    def _aligned_hf_context(self, close_prices: list[float]) -> list[float]:
        context = close_prices[-self.max_context_length :]
        usable_length = len(context)

        if self.patch_length > 0:
            usable_length = (usable_length // self.patch_length) * self.patch_length

        if usable_length < self.min_required_points:
            raise RuntimeError(
                f"TimesFM usable context length {usable_length} is below "
                f"TIMESFM_MIN_REQUIRED_POINTS={self.min_required_points}"
            )

        return context[-usable_length:]

    def _timesfm_confidence(self, outputs: Any, dense_mean: list[float]) -> list[float]:
        full_predictions = getattr(outputs, "full_predictions", None)
        if full_predictions is None:
            return [self.confidence_floor for _ in dense_mean]

        quantiles = full_predictions[0].detach().cpu()
        confidence: list[float] = []
        for idx, pred in enumerate(dense_mean):
            if quantiles.ndim != 2 or idx >= quantiles.shape[0]:
                confidence.append(self.confidence_floor)
                continue
            lower = float(quantiles[idx][0])
            upper = float(quantiles[idx][-1])
            band = abs(upper - lower) / max(abs(pred), 1e-6)
            raw = 1.0 / (1.0 + band)
            confidence.append(max(self.confidence_floor, min(self.confidence_ceiling, raw)))
        return confidence

    def _validate_request(self, payload: PredictRequest) -> None:
        if payload.context_length > self.max_context_length:
            raise ValueError(
                f"context_length {payload.context_length} exceeds "
                f"TIMESFM_MAX_CONTEXT_LENGTH={self.max_context_length}"
            )
        if payload.context_length > len(payload.close_prices):
            raise ValueError("context_length must not exceed len(close_prices)")
        if len(payload.close_prices) < self.min_required_points:
            raise ValueError(
                f"at least {self.min_required_points} close prices are required "
                "for TimesFM stability"
            )
        if any(step > self.max_horizon_step for step in payload.horizons):
            raise ValueError(
                f"horizons contain values above TIMESFM_MAX_HORIZON_STEP={self.max_horizon_step}"
            )

    def _predict_with_baseline(

        self, close_prices: list[float], horizons: list[int]

    ) -> list[PredictionItem]:
        last_price = close_prices[-1]
        short_window = close_prices[-min(8, len(close_prices)) :]
        long_window = close_prices[-min(32, len(close_prices)) :]

        short_mean = mean(short_window)
        long_mean = mean(long_window)
        momentum = 0.0 if short_mean == 0 else (last_price - short_mean) / short_mean
        regime_bias = 0.0 if long_mean == 0 else (short_mean - long_mean) / long_mean

        predictions: list[PredictionItem] = []
        for step in horizons:
            damped_step = math.log(step + 1.0)
            expected_return = momentum * 0.55 + regime_bias * 0.45
            expected_return *= min(1.0, damped_step / 3.5)

            pred_price = max(0.00000001, last_price * (1.0 + expected_return))
            confidence = self._baseline_confidence(close_prices, step, abs(expected_return))
            predictions.append(
                PredictionItem(
                    step=step,
                    pred_price=round(pred_price, 8),
                    pred_confidence=round(confidence, 4),
                )
            )
        return predictions

    def _baseline_confidence(

        self, close_prices: list[float], step: int, expected_move_abs: float

    ) -> float:
        if len(close_prices) < 3:
            return self.confidence_floor

        changes: list[float] = []
        for previous, current in zip(close_prices[:-1], close_prices[1:]):
            if previous <= 0:
                continue
            changes.append(abs((current - previous) / previous))

        realized_vol = mean(changes[-min(32, len(changes)) :]) if changes else 0.0
        signal_to_noise = expected_move_abs / (realized_vol + 1e-9)
        horizon_decay = 1.0 / (1.0 + math.log(step + 1.0))
        raw = 0.25 + min(signal_to_noise, 2.0) * 0.25 + horizon_decay * 0.35
        return max(self.confidence_floor, min(self.confidence_ceiling, raw))

    def describe_runtime(self) -> dict[str, Any]:
        return {
            "model_id": self.model_id,
            "model_name": self.model_name,
            "backend": self.backend,
            "device": self.device,
            "ready": self.ready,
            "load_error": self.load_error,
            "max_context_length": self.max_context_length,
            "max_horizon_step": self.max_horizon_step,
            "min_required_points": self.min_required_points,
            "patch_length": self.patch_length,
        }