<forecast_context>
<dataset>
- Observations: 100
- Observations pooled across series: 200
- Series: 2
- Frequency: D
- Date range: 2023-01-01 to 2023-04-10
- Target: value
- Exogenous columns: none
- Target statistics (store_a): min 0, max 99, mean 49.5, std 29.01
- Target statistics (store_b): min 50, max 149, mean 99.5, std 29.01
- Missing values: none
- Index irregularities: none detected
</dataset>
<profile_decision>
The dataset contains 2 series, so a multi-series forecaster (ForecasterRecursiveMultiSeries) is recommended. Data: 200 observations pooled across 2 series, 'D' frequency. Alternative forecasters: ['ForecasterFoundation']. Estimator: Ridge. A linear model is preferred because the dataset is small (200 observations < 250); gradient boosting is offered as an alternative once more data is available. Alternative estimators: ['RandomForestRegressor', 'LGBMRegressor'].
- Significant lags for store_a (partial autocorrelation, strongest first): 1
- Significant lags for store_b (partial autocorrelation, strongest first): 1
- Suggested window features: mean(window=3), std(window=3), mean(window=7), mean(window=21)
- Suggested calendar features: day_of_week, weekend, month
</profile_decision>
<forecast_plan>
- Forecaster: ForecasterFoundation
- Steps: 3
- Estimator: autogluon/chronos-2-small
- Prediction interval: [0.1, 0.9] (80% coverage)
- Interval method: native
- Primary metric: mean_absolute_scaled_error
- Chosen by the user instead of the rules: forecaster
- Plan: ForecasterFoundation + autogluon/chronos-2-small. No lag or window features: the foundation model forecasts directly from the raw context window. Prediction intervals via native. MASE is scale-independent, enabling fair comparison across differently-scaled series. The model reads up to 8192 observations of each series as context, so the whole history is used (the longest series has 100).

Note: A validated Python script implementing this plan is generated separately. Do not generate code yourself.
</forecast_plan>
<backtesting_strategy>
- steps: 5
- initial_train_size: 70
- gap: 0
- n_folds: 6
- inference_windows: up to 12
</backtesting_strategy>
<deterministic_summary>
Backtested with 6 folds of 5 steps each, starting from an initial training window of 70 observations, without refitting.
</deterministic_summary>
<evaluation_metrics>
 series  MAE   MSE  MASE
store_a  2.5  9.25   0.8
store_b  3.5 16.50   1.2
</evaluation_metrics>
<predictions>
Total rows: 40. Only the first 5 and last 5 rows are shown; the 30 interior rows were not provided. Do not describe trends, growth, or progression across the horizon from these rows, and do not compare an early row against a late row as if they were adjacent. Use the per-column summary below for any statement about the full set of rows.

              level  fold  q_0.1   q_0.5   q_0.9
2023-04-11  store_a     0  900.0  1000.0  1100.0
2023-04-11  store_b     0  902.5  1002.5  1102.5
2023-04-12  store_a     0  905.0  1005.0  1105.0
2023-04-12  store_b     0  907.5  1007.5  1107.5
2023-04-13  store_a     0  910.0  1010.0  1110.0
... (30 rows omitted) ...
2023-04-28  store_b  3  987.5  1087.5  1187.5
2023-04-29  store_a  3  990.0  1090.0  1190.0
2023-04-29  store_b  3  992.5  1092.5  1192.5
2023-04-30  store_a  3  995.0  1095.0  1195.0
2023-04-30  store_b  3  997.5  1097.5  1197.5

Per-column summary (all rows):
  q_0.1: min=900.0, max=997.5, mean=948.75
  q_0.5: min=1000.0, max=1097.5, mean=1048.75
  q_0.9: min=1100.0, max=1197.5, mean=1148.75
Per-level summary of q_0.5 (all rows):
  store_a: min=1000.0, max=1095.0, mean=1047.5
  store_b: min=1002.5, max=1097.5, mean=1050.0
</predictions>
</forecast_context>
