zephyrdn/zephyr-forge-multi-exchange-ohlcv

Zephyr Forge — Multi-Exchange Crypto OHLCV Dataset (Niche Chains v2)

English | 中文

---

English

Cross-validated, multi-exchange OHLCV data for four emerging L1/L2 chains, with full provenance and audit trails.

This is not another raw exchange export. Zephyr Forge is a structured time-series dataset built through a full production pipeline: multi-exchange price alignment, three categories of automated anomaly flagging, a zero-PII compliance declaration, and a complete provenance report for every data combination — built for on-chain/off-chain quant research, LLM financial time-series training, and backtesting systems.

Why this dataset is different

Most free or low-cost crypto OHLCV datasets are a raw export from a single exchange. That creates two real problems: single-exchange spikes, outages, and slippage get baked directly into your model or strategy, and there is no way to verify the data is actually clean. Zephyr Forge addresses both with a full production pipeline:

Feature	Zephyr Forge	Typical free datasets
Multi-exchange cross-validation	✅ Median-based reference price across 4 exchanges (Binance / OKX / Bybit / Coinbase); deviations >3% auto-flagged	❌ Usually single-source
Anomaly detection	✅ Three categories — spike (5σ rolling z-score), zero-gap (outage), stuck (6 consecutive identical closes) — each flagged and never silently deleted, preserving a full audit trail	❌ None, or outliers silently dropped
Zero-PII compliance declaration	✅ Collected exclusively from official public exchange APIs and public chain RPCs; no wallet addresses, user identity, or social data	❌ Usually undeclared
Provenance	✅ Each (chain, pair, timeframe) combination ships with a provenance report covering the full historical range (source APIs, row counts, anomaly stats, date range)	❌ None
Schema versioning	✅ schema_version field guarantees the dataset is replayable and backward-compatible	Inconsistent

Dataset scale

Dimension	Details
Chains	Sui, Aptos, Arbitrum, Optimism
Pairs	SUI/USDT, APT/USDT, ARB/USDT, OP/USDT
Exchanges	Binance, OKX, Bybit, Coinbase (cross-aligned)
Timeframes	1h (base collection) → 4h / 1d (standard resampled aggregates)
Date range	From each chain's listing date (Aptos/Optimism: 2023-01-01, Arbitrum: 2023-03-23, Sui: 2023-05-03) through 2026-07-16, updated on an ongoing basis
Total rows	421,688 (after data-quality cleanup — see "Data quality" below)
File layout	12 (chain × pair × timeframe) combinations, merged into year-partitioned Parquet files — 48 files total
Archive size	~14.8 MB

Data quality (full disclosure)

We treat honest disclosure as part of the product itself.

• The optimism/OP_USDT/4h combination previously contained 230 invalid records (OHLC fields and source fields entirely null — a leftover from a one-off debugging script, not a normal collection artifact). This was roughly 0.74% of that combination's row count. It was removed from the source data after a full SHA256-verified backup, and a defensive filter was added to the export pipeline to prevent similar issues from reaching future releases.
• is_outlier trigger rates differ by chain (some chains show very few or zero flagged anomalies in certain timeframes) — this reflects each chain's actual market volatility, not a failure of the detection logic.
• Start dates differ across chains because the underlying tokens listed at different times — this is a market fact, not a collection gap.

About the free sample

The files under sample/ are not a naive head(N) — they are purposefully stratified so you can verify our core claims before buying the full dataset:

• Covers all 4 chains × 3 timeframes, 600 rows total
• Within each combination, records with is_outlier=True are prioritized (up to 30% of the sample), with the remainder drawn at evenly spaced time intervals so you see the full historical span, not just one slice
• Includes source_manifest_sample.jsonl, letting you verify each row's collection timestamp (ingestion_ts) and official API source (source_url) line by line

When reviewing the sample, pay particular attention to is_outlier, outlier_reason, ref_close, and price_deviation — this is where our core differentiation lives.

Field reference

Field	Description
chain / symbol / exchange / timeframe / ts	Business keys: chain, trading pair, exchange, timeframe, candle open time (UTC)
open / high / low / close / volume / quote_volume	Standard OHLCV fields
clean_status	Processing stage: raw → aligned → filtered → final
is_outlier / outlier_reason	Whether flagged, and type (spike, zero_gap, stuck, multi_exchange_deviation)
ref_close / price_deviation	Multi-exchange reference price (median) and this record's deviation from it
ingestion_ts / source_url	Collection timestamp and original API source, for row-level provenance
schema_version	Currently 1.0; field changes bump the version, keeping historical data replayable

License and how to get the full dataset

• Sample data (this page): free to download, for evaluating data quality and format
• Full dataset: https://zephyrdn.com/products/zephyr-forge — select a plan, complete payment via the crypto checkout, and you'll receive an email within minutes containing a time-limited download link (default 48-hour validity)

Compliance summary

This dataset is collected exclusively from official public sources: exchange REST/WS APIs (OHLCV, volume) and public chain RPCs. It does not collect, store, or deliver any wallet addresses or holdings/transfer activity, user identity information (KYC/email/IP), social media or community data, or any data that could be linked to an individual's on-chain activity. The full declaration is included in the archive at compliance/NO_PII_DECLARATION.md; each data combination ships with compliance/provenance_report_*.md, covering the complete historical range's sources, row counts, and anomaly statistics.

Use cases

• Crypto price forecasting / LLM financial time-series model training
• Quant strategy backtesting (multi-exchange alignment meaningfully reduces single-exchange noise in backtest results)
• On-chain/off-chain research on emerging L1/L2 ecosystems
• Historical analysis of inter-exchange spreads and arbitrage opportunities

License & Legal Compliance

1. Dataset license (derivative work)

The overall structure, cleaning scripts and logic, automated outlier detection algorithms (Spike, Zero-Gap, Stuck, Multi-Exchange Deviation), resampled aggregations, and metadata contained in this dataset (Zephyr Forge Multi-Exchange Crypto OHLCV Dataset) constitute original derived works created by the author.

The structural design, metadata, anomaly labels, and derived statistical indicators of this dataset are licensed under the CC BY-NC-ND 4.0 license. Upon purchasing the complete dataset, you are granted a non-exclusive, non-transferable commercial license for the following scope:

• Internal Commercial R&D: Usage within your organization for quantitative strategy backtesting, risk control modeling, and automated trading system design.
• AI / LLM Model Training: Usage for training or fine-tuning financial time-series prediction models and machine learning algorithms for internal institutional or personal use.
• Prohibition of Redistribution: Redistribution, sublicensing, reselling, or public sharing of the raw Parquet files, raw OHLCV sequences, or un-modified chunks of this dataset—whether paid or free—is strictly prohibited.

2. Exchange Terms of Service & Market Data Compliance

This dataset collects and processes public historical OHLCV market data via official REST/WS APIs from multiple cryptocurrency exchanges. To ensure legal compliance, our review of the Terms of Service (ToS) and Market Data policies for each upstream exchange is detailed below:

① Binance

• Terms Reference: Binance Terms of Use & Binance API Documentation
• Compliance Statement: Data from Binance is collected strictly via its public REST API in full compliance with standard API rate limits without bypassing any authentication mechanisms. All underlying intellectual property and trademark rights belong to Binance and its licensors.

② Coinbase

• Terms Reference: Coinbase Market Data Terms of Use
• Compliance Statement: Coinbase Market Data Terms prohibit unauthorized direct redistribution or commercial ingestion of raw order books and raw market feeds for AI training. This dataset does NOT include or deliver any raw Coinbase order books or un-processed raw market feeds. Public market data from Coinbase is utilized solely as a stateless parameter to compute cross-exchange median reference prices (ref_close). The resulting output represents Derived Statistical Works after multi-source fusion and noise filtering.

③ OKX

• Terms Reference: OKX Terms of Service
• Compliance Statement: Data is collected from OKX official public API endpoints adhering to standard access guidelines. This dataset contains only public, aggregated OHLCV time-series and does NOT contain any proprietary user accounts, trade histories, or PII (Personally Identifiable Information). Original data rights belong to OKX.

④ Bybit

• Terms Reference: Bybit API Terms & Conditions
• Compliance Statement: Access to Bybit public endpoints strictly complies with official Open API rate-limiting rules. Bybit market data serves solely as one of several input streams for multi-exchange cross-validation and resampled aggregations. This heavily modified derived product does not constitute a direct mirror or redistribution of Bybit's raw market data feeds.

3. Disclaimer

1. No Financial Advice: This dataset is provided solely for research, quantitative analysis, and model training purposes. It does not constitute financial, investment, or trading advice. The author assumes no liability for direct or indirect losses incurred from live trading or investment decisions based on this dataset.
2. Data Accuracy: Although anomaly detection and cross-exchange median alignment have been applied, due to extreme cryptocurrency market volatility and potential retroactive exchange API adjustments, 100% error-free data cannot be guaranteed. Users must perform independent verification before live deployment.
3. Trademarks & Copyrights: Binance, OKX, Bybit, Coinbase, Sui, Aptos, Arbitrum, Optimism, and other referenced names are trademarks of their respective owners. The inclusion of these names is solely for descriptive purposes to identify data origins and underlying assets, and does not imply any sponsorship, endorsement, authorization, or affiliation by the respective exchanges or blockchain foundations.

Contact and feedback

If you find a data quality issue, or have a feature suggestion, please reach out at yakovxia@gmail.com. Confirmed issues will be logged into the next version's provenance report.

---

中文

English | 中文

四条新兴公链、四所交易所交叉验证、全链路可溯源的加密货币 K 线数据集。

不是又一份行情导出文件——这是一套经过多所对齐去噪、三类异常自动标记、零 PII 合规声明、逐批次溯源报告齐全的结构化时序数据集，为链上/链下量化研究、LLM 金融时序训练、回测系统设计。

为什么这份数据集不一样

市面上大多数免费/廉价的加密 OHLCV 数据集，只是把某一个交易所的 API 原样导出。这带来两个真实存在的问题：单一交易所的插针、宕机、价格滑点会被原样带进你的模型或策略里；数据来源不可验证，你没法证明这份数据是"干净"的。

Zephyr Forge 用一套完整的生产管道解决这两个问题：

特性	Zephyr Forge	常见免费数据集
多交易所交叉验证	✅ 4 所（Binance / OKX / Bybit / Coinbase）中位数基准对齐，偏离 3% 自动标记	❌ 通常单一来源
异常检测	✅ 插针（5σ 滚动 z-score）/ 宕机（零值）/ 卡死（连续6根K线收盘价相同），三类分别标记且不删除，保留完整审计痕迹	❌ 无，或直接静默删除异常点
零 PII 合规声明	✅ 仅采集交易所官方公开 API + 公链官方 RPC，不含任何钱包地址、用户身份、社交数据	❌ 通常未声明
溯源机制	✅ 每个 (链, 交易对, 时间帧) 组合附完整历史区间的溯源报告（来源 API、行数、异常统计、时间范围）	❌ 无
Schema 版本化	✅ schema_version 字段，保证数据可重放、向后兼容	参差不齐

数据规模

维度	内容
链	Sui、Aptos、Arbitrum、Optimism
交易对	SUI/USDT、APT/USDT、ARB/USDT、OP/USDT
交易所	Binance、OKX、Bybit、Coinbase（多所交叉对齐）
时间帧	1h（基础采集）→ 4h / 1d（标准聚合重采样）
历史区间	各链自上线时起（Aptos/Optimism 2023-01-01、Arbitrum 2023-03-23、Sui 2023-05-03）至 2026-07-16，持续更新
总行数	421,688 行（已完成脏数据清理，详见下方"数据质量说明"）
文件组织	12 个 (链×交易对×时间帧) 组合，按年切分 Parquet 文件，共 48 个文件
压缩包大小	约 14.8 MB

数据质量说明（诚实披露）

我们相信"诚实披露"本身就是这份数据集的一部分价值。以下是我们在生产过程中发现并处理的问题：

• optimism/OP_USDT/4h 组合曾存在 230 条无效记录（OHLC 四要素与来源字段全为空值，来自一次调试脚本产生的历史残留，非正常采集流程产物）。这批记录占该组合总量约 0.74%，已在带 SHA256 校验的完整备份基础上从源数据中清除，并在导出管道中加入了防御性过滤，防止同类问题再次进入交付数据。
• 各链的 is_outlier 触发率不同（例如部分链在某些时间帧下异常记录极少甚至为零），这反映的是各链实际市场波动特征，不代表异常检测机制在该链上失效。
• 由于代币各自的上线时间不同，各链数据起始日期不一致，这是市场事实，不是采集缺陷。

免费样本说明

sample/ 目录下的样本不是简单截取前 N 行，而是按以下策略专门抽样，让你能在下载完整数据前就验证我们的核心卖点：

• 覆盖全部 4 条链 × 3 个时间帧，共 600 行
• 每个组合中，is_outlier=True 的异常标记记录优先抽样（上限 30%），其余按时间均匀间隔抽样，让你能看到跨越完整历史区间的样本，而不只是某个片段
• 附带 source_manifest_sample.jsonl，可逐行核对每条记录的采集时刻（ingestion_ts）与官方 API 来源（source_url）

打开样本后建议重点看：is_outlier / outlier_reason / ref_close / price_deviation 这几列，这是我们和普通数据集的核心差异所在。

数据字段说明

字段	说明
chain / symbol / exchange / timeframe / ts	业务主键：链、交易对、交易所、时间帧、K线开盘时间（UTC）
open / high / low / close / volume / quote_volume	OHLCV 标准字段
clean_status	数据处理阶段：raw → aligned → filtered → final
is_outlier / outlier_reason	是否异常、异常类型（spike 插针 / zero_gap 宕机 / stuck 卡死 / multi_exchange_deviation 多所偏离）
ref_close / price_deviation	多所对齐基准价（中位数）及本条记录相对基准的偏离幅度
ingestion_ts / source_url	采集入库时刻、原始 API 来源，用于逐条溯源
schema_version	当前为 1.0，字段变更会升版本号，保证历史数据可重放

授权与获取方式

• 样本数据（本页）：可直接下载，用于评估数据质量与格式是否符合你的需求
• 完整数据集：https://zephyrdn.com/products/zephyr-forge — 选择套餐后跳转至加密货币支付页面，付款确认后系统会在数分钟内自动发送一封包含限时下载链接的邮件（链接默认 48 小时内有效）

合规声明摘要

本数据集仅采集以下官方公开数据源：交易所官方 REST/WS API（OHLCV、成交量）、公链官方公开 RPC。不采集、不存储、不交付任何钱包地址及持仓/转账行为、用户身份信息（KYC/邮箱/IP）、社交媒体/社区数据，或任何可关联到个人的链上行为数据。完整声明见压缩包内 compliance/NO_PII_DECLARATION.md；每个数据组合附带的 compliance/provenance_report_*.md 提供覆盖完整历史区间的来源、行数、异常统计等溯源信息。

使用场景

• 加密货币价格预测 / LLM 金融时序模型训练
• 量化交易策略回测（多所对齐后的数据可显著降低单一交易所噪声对回测结果的干扰）
• 新兴公链生态的链上/链下关联研究
• 交易所间价差、套利机会的历史分析

## 授权说明（License）

1. 数据集衍生作品许可（Dataset License）

本数据集（Zephyr Forge Multi-Exchange Crypto OHLCV Dataset）包含的清洗代码逻辑、三类异常标记算法（Spike / Zero-gap / Stuck / Multi-exchange deviation）、数据重采样（Resampled Aggregations）以及结构化元数据，均属于作者的原创衍生成果。

针对本数据集中的结构设计、元数据、清洗标记及衍生统计指标，按照 CC BY-NC-ND 4.0 协议授权。购买完整数据集后，您获得非独占、不可转让的商业使用授权，许可范围包括：

• 内部商业研发：用于您或您所在机构内部的量化策略回测、风控模型搭建、自动化交易系统测试。
• AI/LLM 模型训练：允许用于训练、微调（Fine-tune）您或您所在机构内部使用的金融时序预测大模型及机器学习算法。
• 禁止原始再分发与转售：不得将本数据集的原始文件、切分 Parquet 文件通过公开或私密渠道重新打包、出租、直接转售或开源再分发。

---

2. 上游交易所行情数据服务条款与合规声明（Exchange ToS & Market Data Compliance）

本数据集处理并融合了来自多个加密货币交易所公开 REST/WS API 的历史 OHLCV 行情。为了保障商业合规性，本数据集对各交易所的底层服务条款（ToS）及行情数据政策核查与声明如下：

① Binance (币安)

• 条款参考：Binance Terms of Use 及 Binance API Documentation
• 合规声明：本数据集对 Binance 数据的获取完全基于其官方公开 API 接口，并严格遵守标准请求速率限制（Rate Limits），未绕过任何鉴权机制或访问受限数据。数据原始产权及商标权归 Binance 及其许可方所有。

② Coinbase

• 条款参考：Coinbase Market Data Terms of Use
• 合规声明：根据 Coinbase Market Data Terms，Coinbase 明确禁止对其原始 Market Data Feed 进行未经授权的商业再分发或直接用于 AI 模型训练。本数据集不包含、亦未直接交付任何 Coinbase 的原始订单簿或未经加工的原始数据流。 Coinbase 的公共行情仅在数据生产阶段被用于多所中位数基准计算（ref_close），输出的数据属于经过实质性多源融合与滤波后的衍生统计作品（Derived Works）。

③ OKX (欧易)

• 条款参考：OKX Terms of Service
• 合规声明：本数据集采集自 OKX 官方公开 API Endpoint，采集过程符合其接口使用规范。本数据集仅包含公开的市场 OHLCV 汇总数据，不包含任何 OKX 用户的账户数据、交易隐私或私有信息。数据原始产权归 OKX 所有。

④ Bybit

• 条款参考：Bybit API Terms & Conditions
• 合规声明：本数据集对 Bybit 行情数据的调取严格遵循 Open API 的访问规范。数据仅作为多所交叉对齐（Multi-Exchange Alignment）的数据源之一参与聚合计算，并经过了去噪与二次重采样处理，不构成对 Bybit 原始行情数据流（Market Data Feed）的镜像再分发。

---

3. 免责声明（Disclaimer）

1. 无金融与投资建议：本数据集仅供量化研究、数据分析及模型训练使用，不构成任何形式的投资建议、财务建议或交易推荐。作者不承担任何因使用本数据集进行实盘交易或投资决策而产生的直接或间接损失。
2. 数据准确性与完整性：虽然本数据集通过多所交叉对齐和自动检测标记了插针、宕机等异常，但由于加密货币市场的极高波动性及交易所 API 历史数据可能存在的修正，作者不保证数据集 100% 绝对无误。用户应在实盘部署前自行进行数据验证。
3. 商标与知识产权：Binance, OKX, Bybit, Coinbase 以及 Sui, Aptos, Arbitrum, Optimism 等均为其各自权利人的注册商标。本数据集引用上述名称仅用于客观标识数据来源与资产标的，不代表上述平台对本数据集有任何形式的赞助、背书、授权或官方合作关系。

联系方式与反馈

如果你在数据中发现任何质量问题，或对数据集有功能性建议，欢迎通过 yakovxia@gmail.com 与我们联系。我们会将确认的问题记录进下一版本的溯源报告中。