Japan Anime Figure Demand Feature Vectors

Fixed-dimension (48), normalized feature vectors for 650+ Japan anime/figure SKUs — purpose-built for AI agents, ML training pipelines and quant research.

Derived from the multi-shop price snapshot published by japan-anime-figure-price-data. No scraping happens at run time: this Actor re-projects an already-collected dataset into a machine-consumable feature matrix.

What this is not: it is not a back-tested price forecast. See Limitations before you build on it.

---

Quick start

npx @apify/apify-cli run atushi/japan-anime-figure-demand-features --input '{
  "hasPriceOnly": true,
  "minDemandScore": 0.35,
  "limit": 50
}'

Call it from Python

from apify_client import ApifyClient

client = ApifyClient("YOUR_APIFY_TOKEN")
run = client.actor("japan-anime-figure-demand-features").call(
    run_input={"hasPriceOnly": True, "minDemandScore": 0.35, "limit": 50}
)
for item in client.dataset(run["defaultDatasetId"]).iterate_items():
    print(item["figureId"], item["demandScore"], item["vector"][:5])

Call it over HTTP

curl -s "https://api.apify.com/v2/acts/atushi~japan-anime-figure-demand-features/run-sync-get-dataset-items?token=$APIFY_TOKEN" \
  -H 'Content-Type: application/json' \
  -d '{"hasPriceOnly": true, "limit": 10}' | jq '.[0] | {figureId, demandScore, featureDim}'

---

Output shape

Every dataset item:

Field	Type	Meaning
schemaVersion	string	Feature-vector schema version (semver)
featureDim	int	Length of vector; always 48 today
figureId	string	Stable upstream SKU id
name, series, manufacturer, category, janCode	string	Identity
demandScore	number	Composite demand proxy in [0,1]
history	object	points, firstSeen, lastSeen, velocityAvailable
coverage	object	hasOfferDetail, offerDetailRows, hasAggregatePrice, hasMsrp, hasJan
features	object	48 named features (canonical order)
featureNames	array	Ordered names; featureNames[i] ↔ vector[i]
vector	array	48 floats in featureNames order

The full schema manifest ships in the run's key-value store as FEATURE_MANIFEST (dimension, ordered names, per-dimension documentation and normalization constants). Fetch it once and build your column mapping from it instead of hard-coding indices.

Feature groups

Group	Count	Examples
Price	8	price_min_norm, msrp_norm, discount_depth, premium_ratio
Supply	8	stock_ratio, offer_count_norm, shop_diversity_ratio
Cross-shop dispersion	4	shop_price_cv, cheapest_shop_advantage, instock_price_premium
Temporal	8	age_days_norm, release_year_norm, velocity_7d, history_available
Metadata completeness	8	has_msrp, has_jan, confidence
Categorical / identity	8	is_scale_figure, manufacturer_pop_norm, hash-bucket embeddings
Relative price rank	3	price_rank_in_series, price_rank_in_category, price_rank_in_manufacturer
Composite	1	demand_score

Normalization: every component lies in [0,1], except instock_price_premium and velocity_* which are clipped to [-1,1]. Absolute prices are divided by 50,000 JPY and offer counts by 30 — both constants are exposed in the manifest.

demand_score

demand_score = 0.35 * stock_ratio
             + 0.25 * log_offer_count_norm
             + 0.20 * discount_depth
             + 0.10 * (1 - age_days_norm)
             + 0.10 * (1 - price_spread_ratio)

A supply-side demand-pressure proxy. It is not a forecast and has not been back-tested against realised sales.

---

Limitations

• Single collection point. The shipped upstream snapshot was captured on one date, so history.points is 1 for the 132 rows that carry offer detail and 0 for the rest. velocity_7d / velocity_30d are therefore emitted as 0.0. Gate on history_available == 1.0 before using velocity features; where it is 0.0 the values carry no information.
• Aggregate-only rows. 181 upstream rows advertise a non-zero offer count (and often a price) while shipping an empty offers[] array. Price features still resolve from the aggregate figures, but every per-shop feature (shop_price_cv, shop_price_range_ratio, cheapest_shop_advantage, distinct_shop_norm, shop_diversity_ratio, instock_price_premium) is 0.0 for them — read coverage.hasOfferDetail first, or pass requireOfferDetail: true.
• Sparse price coverage. 292 of 654 upstream rows carry a price; the rest have metadata only. Use hasPriceOnly: true to exclude them.
• Hash-bucket embeddings. manufacturer_bucket_norm and friends are deterministic SHA-256 buckets, not trained embeddings. They give models a stable categorical signal, not semantics.
• Japan-market scope. Figures, hobby goods and their secondary market only.

Pricing

Pay-Per-Event, $0.002 per dataset item returned — the account-standard price point shared by 46 of 73 portfolio Actors. A one-time Actor-start event applies per run ($0.00005 per GB of memory). No subscription, no minimum.

Input reference

Field	Type	Default	Notes
figureIds	string[]	–	Restrict to specific upstream SKUs
series / character / manufacturer	string	–	Case-insensitive partial match
categories	string[]	–	Scale Figure, Anime Figure
minDemandScore / maxDemandScore	number	–	Filter on demandScore
hasPriceOnly	bool	false	Drop rows with no observed price
requireOfferDetail	bool	false	Drop rows with an empty offers[]
includeFeatures	bool	true	Attach named feature dict
includeVector	bool	true	Attach ordered vector
limit / offset	int	100 / 0	Ordered by descending demandScore, then figureId

Provenance

Upstream dataset: japan-anime-figure-price-data (MyFigureList primary, 23 distinct shops observed, JAN/GTIN keyed where available). Each record carries sourcesMerged and confidence from the upstream collector via provenance in the manifest pipeline.

License

MIT.