CWM — Consumer World Model · 消費者理解 · コア

CWMとは？消費者行動を理解するモデル。

行動と消費者シグナルを分析し、類似顧客を発見。ワークスペース、API、MCP から利用できます。

新規 CWM アカウントに 500,000 クレジット

Loop Marketingアプリケーション Harnessランタイム CWMConsumer World Modelコア

Next behaviors, purchase windows and demand context

One member prediction returns next behaviors and personal purchase windows, combining consumer signals with brand demand context when the tenant fusion head is ready.

Interactive example

One consumer. One prediction response.

All states and numbers are synthetic examples, not live predictions or workspace readiness. No customer data is sent.

Choose an example state

Awaiting trainingValidated exampleMissing permissionInsufficient history

Partial result

Behavior predictions remain available. Purchase windows stay unavailable until the brand's fusion head is trained and validated.

What could happen next?

One final distribution over this example's four supported events. Next-event probabilities are uncalibrated.

Next event	Probability
Purchase	37%
Portal visit	28%
Email click	21%
Coupon redemption	14%

Why is an event missing?

Email sent is an input context event, not a prediction target. Its probability is null. Other unsupported or uncovered events also stay unavailable.

Will they buy within…?

Personal CWM signals + TimesFM brand demand context.

7 days: Unavailable
30 days: Unavailable
90 days: Unavailable

Chance of at least one purchase. These windows are cumulative; do not add them together or to the event distribution.

Use the same result in your workflow

One member request returns distribution and purchaseWindows together.

Explore API / MCP calls

curl --request GET \
  "https://YOUR_API_HOST/api/v2/cwm/members/YOUR_MEMBER_UUID/next-behaviors" \
  --header "Authorization: Bearer YOUR_API_KEY"

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "cwm_predict_next_behaviors",
    "arguments": {
      "member_id": "YOUR_MEMBER_UUID"
    }
  }
}

Replace the API host, member UUID and API key with your workspace values.

Required access · predictions:read returns behaviors. Also grant metrics:read (or *) to enable purchase windows when the model is ready.

View response excerpt

Synthetic excerpt only. Brand-model responses also include the full event catalog and coverage metadata; public-prior references use a separate six-class vocabulary.

{
  "predictionKind": "next_event",
  "unit": "probability_0_1",
  "status": "degraded",
  "distribution": {
    "status": "ok",
    "vocabulary": "member_event",
    "normalization": "conditional_on_supported_events",
    "calibrated": false,
    "rows": [
      {
        "key": "purchase.completed",
        "label": "Purchase",
        "probability": 0.37,
        "rank": 1,
        "availability": "predicted"
      },
      {
        "key": "portal.visited",
        "label": "Portal visit",
        "probability": 0.28,
        "rank": 2,
        "availability": "predicted"
      },
      {
        "key": "email.clicked",
        "label": "Email clicked",
        "probability": 0.21,
        "rank": 3,
        "availability": "predicted"
      },
      {
        "key": "coupon.redeemed",
        "label": "Coupon redeemed",
        "probability": 0.14,
        "rank": 4,
        "availability": "predicted"
      },
      {
        "key": "email.sent",
        "label": "Email sent",
        "probability": null,
        "rank": null,
        "availability": "context_only"
      }
    ]
  },
  "purchaseWindows": {
    "unit": "probability_0_1",
    "probabilities": {
      "7": null,
      "30": null,
      "90": null
    },
    "calibrated": false,
    "envelope": {
      "status": "unavailable",
      "backend": "cwm-timesfm-2.5",
      "reasonCodes": [
        "fusion_not_trained"
      ]
    }
  },
  "reasonCodes": [
    "fusion_not_trained"
  ]
}

Try CWM →REST reference →MCP reference →

Synthetic data · TimesFM 2.5

Explore daily demand trends

An example forecast from 128 days of synthetic history. Compare the recent pattern with future daily demand and its quantiles.

Forecast horizon

7 days30 days90 days

Synthetic demand units

Recent daily average: 20.0
Forecast daily average: 20.1
Forecast total: 604.3

History · recent 28 daysPoint forecastShow q10–q90

Inspect a forecast dayDay 1 · 2026-10-05

Point forecast: 22.9
q10: 22.8
q90: 23.0

View daily forecast data

Date (UTC)	Point forecast	q10–q90
2026-10-05	22.9	22.8–23.0
2026-10-06	21.3	21.2–21.4
2026-10-07	18.7	18.6–18.8
2026-10-08	17.1	17.0–17.1
2026-10-09	17.7	17.6–17.7
2026-10-10	20.0	20.0–20.1
2026-10-11	22.3	22.3–22.4
2026-10-12	22.9	22.9–22.9
2026-10-13	21.3	21.2–21.3
2026-10-14	18.7	18.6–18.7
2026-10-15	17.1	17.0–17.2
2026-10-16	17.7	17.6–17.7
2026-10-17	20.0	19.9–20.1
2026-10-18	22.4	22.3–22.4
2026-10-19	22.9	22.8–23.0
2026-10-20	21.3	21.2–21.4
2026-10-21	18.7	18.6–18.7
2026-10-22	17.1	17.0–17.1
2026-10-23	17.7	17.6–17.7
2026-10-24	20.0	19.9–20.0
2026-10-25	22.3	22.3–22.4
2026-10-26	22.9	22.8–22.9
2026-10-27	21.3	21.2–21.3
2026-10-28	18.7	18.6–18.7
2026-10-29	17.1	17.0–17.2
2026-10-30	17.7	17.6–17.7
2026-10-31	20.0	20.0–20.1
2026-11-01	22.4	22.3–22.4
2026-11-02	22.9	22.8–22.9
2026-11-03	21.3	21.2–21.3

Download example data (JSON) ↓

Synthetic example, generated with TimesFM 2.5 weights. The forecast starts on 2026-10-05 (UTC); dates are fixed for the demo. Quantile coverage is uncalibrated. Values do not represent a customer's demand or individual purchase probabilities.

One complete next-behavior distribution

Receive one final probability distribution and ranking. The evaluated brand model takes priority, with up to 23 covered targets. When it is unavailable, internal public-prior models form a single six-class reference. Unsupported and context-only events remain unavailable.

7 / 30 / 90-day purchase windows

Included automatically in the same member response: a tenant-private lightweight head combines CWM signals with TimesFM brand trends. Ready heads serve directly; missing training or evaluation leaves only the purchase windows unavailable.

Daily brand and category forecasts

Where the TimesFM 2.5 service is configured and daily history is complete, forecast revenue, orders or units sold for up to 90 days, with point estimates and nine quantiles.

How to read these results

Next-event probabilities sum to 100% over supported candidates and are uncalibrated. Purchase windows separately estimate at least one purchase within 7/30/90 days and are cumulative, never normalized with event classes. Elapsed silence is not conditioned in the event vector. The public-prior reference has not earned workspace validation.

Demand quantiles describe aggregate uncertainty; they are not individual purchase intervals and have not been calibrated. Missing data remain unavailable. No accuracy lift is asserted before your own evaluation.

Forecast API and MCP reference →