# HarvestGroup360 > HarvestGroup360 (legal name AMII LTD; operational headquarters in Warsaw) builds quantitative data infrastructure: market-data normalisation, ultra-low-latency API pipelines, and institutional-grade reference data. The company publishes market-data-normalizer, an MIT-licensed Python library with no runtime dependencies, and a research blog documenting the specific ways a research pipeline produces a number that is wrong in a direction that flatters it. Open to everyone, from independent developers to funds. ## Open source - [market-data-normalizer on GitHub](https://github.com/Harvestgroup360/market-data-normalizer): the library. Sixty tagged releases, 1,815 tests, type-checked clean, no runtime dependencies. - [market-data-normalizer on PyPI](https://pypi.org/project/market-data-normalizer/): `pip install market-data-normalizer`. Forty-seven published releases. - [ROADMAP.md](https://github.com/Harvestgroup360/market-data-normalizer/blob/main/ROADMAP.md): what exists, what has been asked for, and what the project has decided against and why. ## Key pages - [https://harvestgroup360.com/](https://harvestgroup360.com/): Company overview, products and coverage. - [https://harvestgroup360.com/about](https://harvestgroup360.com/about): The team, the legal entity and what the company does. - [https://harvestgroup360.com/blog](https://harvestgroup360.com/blog): All 72 research articles, newest first. - [https://harvestgroup360.com/api](https://harvestgroup360.com/api): API documentation and integration notes. - [https://harvestgroup360.com/terminal](https://harvestgroup360.com/terminal): The data terminal. - [https://harvestgroup360.com/community](https://harvestgroup360.com/community): Public comments on the library, each quoted in full with a link to its source. - [https://harvestgroup360.com/partnership](https://harvestgroup360.com/partnership): Partnership and investment enquiries. - [https://harvestgroup360.com/contact](https://harvestgroup360.com/contact): Contact details. ## Research articles ### Market microstructure - [The Biggest Trade of the Day Was Not a Trade](https://harvestgroup360.com/article58) (2026-09-06): The closing cross is one print, a tenth of the day's volume and no aggressor. What it does to a VWAP benchmark, an inferred side and a volatility figure. - [Your Volume Profile Already Knows How the Year Ends](https://harvestgroup360.com/article55) (2026-09-02): An intraday profile fitted on the whole sample judges January against a curve that contains December. How much that distorts, measured on both. - [Trade Classification: Working Out Who Crossed the Spread](https://harvestgroup360.com/article42) (2026-08-13): How the tick rule, quote rule and Lee-Ready algorithm infer the aggressor side of a trade, and what order-flow analysis needs it for. - [Microstructure Feature Engineering: Why Your Daily Close Prices Are Lying To You](https://harvestgroup360.com/article31) (2026-07-29): Training on daily close prices discards the microstructure signal inside the session. What is lost, and what to build features from instead. - [Fragmented Liquidity: Intelligent Order Routing Strategies](https://harvestgroup360.com/article15) (2026-06-15): How smart order routing engines split and direct flow across venues to minimise slippage and achieve best execution in fragmented markets. - [The Shift to Direct Market Access](https://harvestgroup360.com/article10) (2026-06-15): Why serious algorithmic trading firms abandon retail brokerages in favor of unadulterated Direct Market Access (DMA). - [Level II Order Book Dynamics: Reading the Tape Programmatically](https://harvestgroup360.com/article7) (2026-06-15): Extracting alpha from L2 data feeds requires sophisticated parsing. Learn how institutional quants analyze order book imbalances in real-time. - [Market Microstructure: The Hidden Mechanics of Liquidity](https://harvestgroup360.com/article1) (2026-06-15): Understanding the order book, tick data, and how liquidity providers interact with taker flow is crucial for latency-sensitive strategies. ### Latency and infrastructure - [We Were Asked to Rewrite It in Rust. First We Measured.](https://harvestgroup360.com/article48) (2026-08-25): Exact Decimal arithmetic cost 3.1x a float loop, not the order of magnitude usually assumed. What an honest benchmark found in our own code instead. - [What HFT Infrastructure Actually Looks Like: The Path of an Order, Layer by Layer](https://harvestgroup360.com/article37) (2026-08-06): What HFT infrastructure actually looks like: colocation, feed handlers, kernel bypass, pre-trade risk and clock sync — the path of an order, layer by layer. - [Python vs C++ in Quant Finance: The Architectural Divide](https://harvestgroup360.com/article34) (2026-08-02): Discover the critical architectural separation in quantitative finance: why Python dominates machine learning research while C++ is essential for execution. - [The True Cost of Latency: Why 1 Millisecond Means Structural Adverse Selection](https://harvestgroup360.com/article33) (2026-07-31): A millisecond of latency is not a speed problem, it is structural adverse selection. What that costs and where the delay actually comes from. - [The Role of FPGA Technology in Ultra-Low Latency Trading](https://harvestgroup360.com/article25) (2026-07-19): How hardware acceleration is revolutionizing tick data processing and algorithmic execution by pushing logic down to the silicon level. - [Execution Quality: Why Infrastructure Trumps Strategy Logic](https://harvestgroup360.com/article12) (2026-06-15): Even the most mathematically sound strategy will bleed capital if deployed on subpar infrastructure. Latency optimization is the true edge. - [Institutional Cross-Connects: Why NY4 and LD4 Matter](https://harvestgroup360.com/article5) (2026-06-15): An inside look at the physical fiber optic cross-connects that power the fastest data transmission between hedge funds and matching engines. - [Optimizing Algorithmic Execution: The Latency Arms Race](https://harvestgroup360.com/article2) (2026-06-15): From colocated servers in NY4 to bypassing the TCP stack, we explore the hardware and software layers required for microsecond execution. ### Connectivity, APIs and reliability - [The Anatomy of a High-Frequency Trading API: Why Milliseconds Matter](https://harvestgroup360.com/article24) (2026-07-18): Exploring the architectural requirements of Tier-1 FIX pipelines and the true cost of latency in quantitative trading. - [Building Enterprise-Grade Fail-Safes in Automated Trading](https://harvestgroup360.com/article20) (2026-07-14): Lessons learned from a Tier-1 security audit: How to protect client capital from black-swan events and infrastructure failures. - [Zero-Trust Architecture in Algorithmic Trading](https://harvestgroup360.com/article17) (2026-06-20): Applying zero-trust security architecture to trading infrastructure: identity, segmentation, and least-privilege access for execution systems. - [Overcoming API Rate Limits in High-Frequency Environments](https://harvestgroup360.com/article13) (2026-06-15): Strategies for connection pooling, efficient data batching, and handling HTTP 429 Too Many Requests in rigorous institutional APIs. - [Designing Fault-Tolerant Trading Algorithms](https://harvestgroup360.com/article6) (2026-06-15): How to implement robust circuit breakers, handle corrupted data packets, and manage failovers in high-frequency trading pipelines. - [FIX Protocol vs. WebSockets: Choosing the Right Pipeline](https://harvestgroup360.com/article3) (2026-06-15): A technical comparison between the Financial Information eXchange (FIX) protocol and modern WebSockets for streaming high-frequency data. ### Backtest validation and overfitting - [Your Best Backtest Is the Luckiest One You Tried](https://harvestgroup360.com/article72) (2026-09-22): Combinatorially symmetric cross-validation measures whether choosing the in-sample winner picks something that keeps working: 0.75 on twenty identical strategies, 0.39 once one genuinely better strategy is added. - [The Sharpe Ratio of a Residual Is Always Zero](https://harvestgroup360.com/article64) (2026-09-14): An intercept absorbs the mean of the residuals by construction, so a Sharpe ratio on them is zero for any strategy. What to measure instead, and a five-year factor decomposition. - [Twelve Defensible Pipelines, One Published Number](https://harvestgroup360.com/article62) (2026-09-12): Three cleaning decisions nobody argued about turn one backtest into twelve, between an annualised Sharpe of 0.49 and 0.71, and the clipping threshold moves the answer more than the other two combined. - [Every Backtest Has a Start Date Somebody Chose](https://harvestgroup360.com/article61) (2026-09-11): One strategy, one dataset, twenty-five start dates. It loses money over the full sample and makes money from four of them, and only one number ever gets published. - [AI Made Strategies Cheap. Evidence Did Not Get Cheaper.](https://harvestgroup360.com/article56) (2026-09-04): What AI actually changed in quantitative trading, what it multiplied, and why the number of strategies you tried is now the hardest figure to obtain. - [Your Backtest Knew the Price Before Your Machine Did](https://harvestgroup360.com/article54) (2026-09-01): A venue timestamp is not an arrival. How to measure the delay between the two, and what keying research on the venue stamp quietly buys a backtest. - [Two Hundred Coin Flips and the One That Looked Like Skill](https://harvestgroup360.com/article46) (2026-08-22): Two hundred strategies made of random numbers. The best scores 2.43 annualised and passes every standard check. What the deflated Sharpe says. - [The Bugs That Make Your Backtest Better Are the Ones Nobody Finds](https://harvestgroup360.com/article44) (2026-08-19): How look-ahead bias enters an as-of join: nearest-observation lookups, bars read on their label instead of their close, and forward-fills that never expire. - [The Overfitting Trap: Why 99% Backtest Accuracy Should Scare You](https://harvestgroup360.com/article38) (2026-08-07): Why stellar backtest accuracy is a warning sign in trading ML: overfitting, data leakage and look-ahead bias — and the validation habits that catch them. - [The Evolution of Statistical Arbitrage in Modern Markets](https://harvestgroup360.com/article19) (2026-07-13): How machine learning and ultra-low latency infrastructure have fundamentally transformed classic pair trading. - [The Fallacy of Retail Backtesting Platforms](https://harvestgroup360.com/article4) (2026-06-15): Why testing algorithms on synthetic liquidity without factoring in true market depth and queue position leads to massive live-execution degradation. ### Market-data quality and normalisation - [A Fat Tail and a Fat Finger Look Identical](https://harvestgroup360.com/article70) (2026-09-20): Why ordinary z-scores miss the contamination they are meant to catch, how a robust scale avoids it, and why clipping a one-sided tail halved a Sharpe ratio instead of raising it. - [Your Price File Is Either Prints or Arithmetic](https://harvestgroup360.com/article52) (2026-08-30): A venue only accepts multiples of a tick, so raw prints sit on the grid. One pass tells you whether your file is prints, mids, VWAPs or adjusted. - [The File That Decides Which Company Your Rows Are About](https://harvestgroup360.com/article47) (2026-08-24): A ticker names different companies at different times. How to build the reference file that records which one, and how to check the file you were given. - [Nine Questions That Find the Errors Nobody Reports](https://harvestgroup360.com/article45) (2026-08-21): Nine checks for look-ahead bias, label leakage, survivorship and revised data — one per stage of a research pipeline, and what a real pass looks like. - [Split-Adjusted Prices: The Crashes in Your Data That Never Happened](https://harvestgroup360.com/article41) (2026-08-12): How splits, dividends and futures rolls create fake returns in historical price data, and how back-adjustment repairs a price series correctly. - [Market Data Normalization: A Practical Guide for Quantitative Teams](https://harvestgroup360.com/article39) (2026-08-08): A practical guide to market data normalization: canonical schemas, deduplication, quality checks, bar sampling, trading sessions and storage formats. - [The Hidden Cost of Dirty Market Data: Why Your Algorithmic Models Bleed in Production](https://harvestgroup360.com/article35) (2026-08-03): Dirty market data quietly wrecks live trading models — bad ticks, gaps and stale quotes. See how data-quality checks stop your edge bleeding in production. - [The Future of Quantitative Data Infrastructure](https://harvestgroup360.com/article14) (2026-06-15): From FPGA hardware acceleration to microwave transmission networks, what the next decade of algorithmic trading infrastructure will look like. - [Building a Scalable Tick Data Architecture](https://harvestgroup360.com/article9) (2026-06-15): Storing and querying petabytes of raw tick data requires specialized time-series databases. We compare kdb+, InfluxDB, and ClickHouse. ### Risk, execution and benchmarks - [A Backtest Reports What the Strategy Earned. An Investor Keeps Less.](https://harvestgroup360.com/article71) (2026-09-21): A fund's headline fee rate is not the share of profit it takes. Crystallisation frequency and the high-water mark moved it from 53 to 83 per cent on identical returns. - [How Often the Book Is Traded Back Is an Assumption](https://harvestgroup360.com/article69) (2026-09-19): The rebalancing frequency in a backtest is an unstated assumption. Turnover rises steadily with it and the return follows no order; a drift band beats a calendar on both control and trading. - [A Return Has to Beat Something](https://harvestgroup360.com/article68) (2026-09-18): Why the hurdle under a Sharpe ratio is a decision, why a constant rate and a rate series are different operations, and the day-count convention that understates a money-market hurdle by 1.39 per cent. - [A Maximum Drawdown Is a Maximum](https://harvestgroup360.com/article67) (2026-09-17): A maximum drawdown is an order statistic, so it grows with the length of the sample. What duration adds that depth leaves out, and why two backtests of different lengths are not comparable. - [The Square Root of Twelve Is an Assumption](https://harvestgroup360.com/article66) (2026-09-16): Multiplying by the square root of the calendar assumes the returns are independent. On a book with partly stale marks that assumption costs 27.55 per cent of the annualised figure. - [The Average Return Is Not a Return Anybody Received](https://harvestgroup360.com/article65) (2026-09-15): The arithmetic mean always exceeds the rate that compounds, the gap grows with volatility, and doubling the leverage quadruples it. Why annualising is the comparison with a guaranteed direction. - [Forty Positions Is Not Forty Bets](https://harvestgroup360.com/article63) (2026-09-13): One eigenvalue carries half the variance of forty names. What the effective bet count is, how it differs from the effective observation count, and what it does to a t statistic of 3.10. - [In a Backtest, Every Stop Fills](https://harvestgroup360.com/article59) (2026-09-07): A halted instrument cannot be sold, but a backtest reading the tape sells anyway. What that exit is worth, and how to count the decisions that were never available. - [The Daily Close Your Model Read at Ten in the Morning](https://harvestgroup360.com/article49) (2026-08-26): A daily close is not knowable at nine in the morning. How much future a label-keyed join leaks, why every grid point is affected, and how to measure it. - [Your Own Trades Are in the Benchmark You Are Measured Against](https://harvestgroup360.com/article43) (2026-08-16): Why VWAP includes your own prints, how that flatters execution scores, and what participation rate and implementation shortfall really measure. - [Building Trust in Institutional Trading Infrastructure](https://harvestgroup360.com/article28) (2026-07-23): Why transparency, uptime reliability, and authentic community feedback are the cornerstones of modern quantitative execution. - [Real-Time Risk Engines: Pre-Trade Validation](https://harvestgroup360.com/article18) (2026-07-01): How pre-trade risk engines enforce position, exposure, and fat-finger checks in microseconds before orders ever reach the exchange. - [Navigating Slippage: Managing Risk During Macroeconomic Spikes](https://harvestgroup360.com/article8) (2026-06-15): How NFP and FOMC announcements instantly drain liquidity pools, and the quantitative models required to avoid catastrophic slippage. ### Machine learning and alternative data - [Out-of-the-box NLP APIs vs. Bare-Metal Micro-Transformers](https://harvestgroup360.com/article30) (2026-07-26): Why standard NLP APIs like ChatGPT and BERT fail in high-frequency trading, and why bare-metal micro-transformers are the only solution for latency. - [Moving from Standard NLP to Financial Data: The Reality of Sentiment Modeling](https://harvestgroup360.com/article27) (2026-07-23): How transitioning from traditional natural language processing to chaotic market data completely changes the paradigm of AI research. - [Deep Learning for Order Book Imbalance: Predictive Signals in Microstructure](https://harvestgroup360.com/article26) (2026-07-21): How neural networks extract high-probability directional signals from the chaotic depths of Level II market data. - [Machine Learning in Quantitative Finance: The New Standard for Institutional Alpha](https://harvestgroup360.com/article23) (2026-07-17): How advanced machine learning architectures are systematically replacing traditional statistical arbitrage models in Tier-1 trading operations. - [Deep Reinforcement Learning in High-Frequency Market Making](https://harvestgroup360.com/article21) (2026-07-15): How neural networks are replacing classic analytical models to dynamically manage inventory risk and capture spreads in non-stationary crypto markets. - [Integrating Alternative Data into High-Frequency Pipelines](https://harvestgroup360.com/article16) (2026-06-15): How high-frequency trading desks turn alternative data and NLP-driven sentiment into structured, low-latency signals for execution models. - [Machine Learning in HFT: Separating Signal from Noise](https://harvestgroup360.com/article11) (2026-06-15): Deploying deep learning models in low-latency environments is notoriously difficult. We discuss the difficulties of overfitting on historical data. ### Engineering practice and the open-source library - [What We Refuse to Build, and the One Refusal We Reversed](https://harvestgroup360.com/article53) (2026-08-31): Five things our open-source library will not do, the reason written next to each, and the one we overturned when the objection turned out to be testable. - [Every Name on Today's List Is a Name That Survived](https://harvestgroup360.com/article50) (2026-08-27): A today-list gets 40 of 100 index names wrong four years back — half dropped, half held early. How to build a membership record that does not flatter. - [market-data-normalizer 1.0: Composable Pipelines, NDJSON I/O and a Zero-Dependency CLI](https://harvestgroup360.com/article36) (2026-08-04): market-data-normalizer 1.0: composable pipelines, NDJSON I/O and a zero-dependency CLI that turn raw multi-venue trade feeds into research-ready OHLCV bars. - [Open-Sourcing market-data-normalizer v0.4.0: Guardrails for Dirty Data](https://harvestgroup360.com/article32) (2026-07-30): market-data-normalizer v0.4.0 adds guardrails for dirty ticks and a VWAP that refuses to be computed from data it should not trust. - [Culture & Engineering: Bridging Data Science and Low-Level Systems](https://harvestgroup360.com/article29) (2026-07-26): Discussing the hiring philosophy at HarvestGroup360 and the critical bridge between mathematical modeling and execution architecture. - [The Future of Market Intelligence: Democratizing Tier-1 Analytics](https://harvestgroup360.com/article22) (2026-07-16): How HarvestGroup360 is bridging the gap between independent quantitative analysts and institutional-grade data infrastructure. ### Further articles - [The File Did Not Change. The Answer Did.](https://harvestgroup360.com/article60) (2026-09-09): The same data and the same code produced a publishable result and an unpublishable one. Only an argument moved, and nothing recorded which was used. - [A Thousand Rows, Two Hundred Observations](https://harvestgroup360.com/article57) (2026-09-05): Overlapping labels inflate every t-statistic computed on them. How to count the observations you actually have, and what it does to a published result. - [The Year Everyone Calls 252 Sessions Long Was 251](https://harvestgroup360.com/article51) (2026-08-29): US equities were open 251 sessions and 97,530 minutes in 2026, not 252 by 390. Where the annualisation constant comes from and what assuming it costs. - [Trading Sessions and Time Zones: The Bug That Comes Back Twice a Year](https://harvestgroup360.com/article40) (2026-08-10): Why market hours break data pipelines twice a year: daylight saving, overnight sessions, trading-day attribution, and how to filter sessions correctly. ## Elsewhere - [https://www.linkedin.com/company/harvestgroup-360](https://www.linkedin.com/company/harvestgroup-360): Company page; releases and engineering notes. - [https://medium.com/@HarvestGroup360](https://medium.com/@HarvestGroup360): Longer write-ups. - [https://github.com/Harvestgroup360](https://github.com/Harvestgroup360): The engineering organisation. - [https://www.crunchbase.com/organization/harvestgroup360](https://www.crunchbase.com/organization/harvestgroup360): Company profile. - [https://wellfound.com/company/harvestgroup360](https://wellfound.com/company/harvestgroup360): Company profile and roles. - [https://www.wikidata.org/wiki/Q140830770](https://www.wikidata.org/wiki/Q140830770): Structured entity record.