Opf Prediction

---
name: opf-prediction
description: "Use NuPIC legacy Online Prediction Framework models, experiment
  directories, prediction outputs, and checkpoints."
disable-model-invocation: true
metadata:
  disco-role: operating
license: MIT
---

# OPF Prediction

Use this sub-skill when the task mentions NuPIC legacy OPF, `ModelFactory`, `HTMPredictionModel`, `model.run`, `result.inferences`, `multiStepBestPredictions`, `run_opf_experiment`, experiment `description.py` directories, or OPF checkpoints.

## Route first

- For CSV header rows, field type conversion rules, stream sources, and model parameter YAML/JSON structure, read [`../data-and-configuration/`](../data-and-configuration/) before changing OPF code.
- For encoder, Spatial Pooler, Temporal Memory, classifier, anomaly-score, or SDR internals below OPF, read [`../htm-algorithms/`](../htm-algorithms/).
- For models generated by a swarm, read [`../swarming/`](../swarming/) first, then return here to run the generated best OPF model.
- For package installation, Python 2.7, `nupic.bindings`, `numpy`, `pycapnp`, or capnproto failures, start with root troubleshooting at [`../../references/troubleshooting.md`](../../references/troubleshooting.md) and then use the OPF-specific troubleshooting below.

NuPIC legacy OPF is a Python 2.7-era API. Expect environments to need the legacy `nupic` package, compiled `nupic.bindings`, `numpy` 1.12.x-era compatibility, and `pycapnp`/capnproto for some serialization paths.

## Fast workflow map

1. **Create and run a model from CSV**: load model params, call `ModelFactory.create(modelConfig, logLevel=40)`, call `model.enableInference({'predictedField': '<field>'})`, convert each CSV value to the declared Python type, call `model.run(record)`, and read `result.inferences`. Full recipe: [`references/opf-workflows.md`](references/opf-workflows.md).
2. **Extract one- or multi-step predictions**: read `result.inferences['multiStepBestPredictions'][step]` and confidence from `result.inferences['multiStepPredictions'][step][predictedValue]`, allowing warm-up `None`/empty predictions. Shapes and API keys: [`references/api-reference.md`](references/api-reference.md).
3. **Run an OPF experiment directory**: use the installed experiment runner against an experiment directory containing `description.py`; list tasks/checkpoints before long runs. Command patterns and checkpoint layout: [`references/opf-workflows.md`](references/opf-workflows.md).
4. **Checkpoint or reload a model**: direct API uses `model.save(saveModelDir)` and `ModelFactory.loadFromCheckpoint(saveModelDir)`; experiment runs use `savedmodels/<label>.nta` under the experiment directory. Details: [`references/api-reference.md`](references/api-reference.md).
5. **Debug failures**: predicted-field mismatches, malformed model params, unconverted raw strings, missing inference keys, runner usage errors, checkpoint directory shape, missing bindings/capnp, and Python 2 print syntax are covered in [`references/troubleshooting.md`](references/troubleshooting.md).

## Bundled scripts

- Run [`scripts/opf_prediction_smoke.py`](scripts/opf_prediction_smoke.py) as a deterministic import/API smoke check; with `--params`, `--csv`, and `--predicted-field` it can run a tiny bounded OPF CSV pass without depending on the original repository checkout.
- Run [`scripts/opf_experiment_help.py`](scripts/opf_experiment_help.py) to print self-contained OPF experiment runner command patterns, or use `--runner-help`/`--check-import` inside an installed NuPIC legacy environment.

## Minimal direct API skeleton

```python
from nupic.frameworks.opf.model_factory import ModelFactory

model = ModelFactory.create(modelConfig, logLevel=40)
model.enableInference({'predictedField': 'consumption'})

record = {'timestamp': timestamp_datetime, 'consumption': 21.2}
result = model.run(record)
best = result.inferences['multiStepBestPredictions']
all_predictions = result.inferences['multiStepPredictions']
if best.get(1) is not None:
    one_step = best[1]
    confidence = all_predictions[1][one_step]
```

Keep OPF records typed before `model.run`; raw CSV strings are a common source of misleading failures.