Data Sourcing and Quality Control
The foundation of any quantitative system is data, but not all data is created equal. In my experience at DONGZHOU LIMITED, I've learned a painful lesson: garbage in, gospel out. No, I meant garbage in, garbage out. But the subtle truth is more dangerous—garbage in, *plausible-looking* garbage out. A model trained on flawed data will produce confident, wrong predictions, and those can be catastrophic. So our first pillar is a robust, multi-layered data ingestion and quality assurance pipeline.
We aggregate data from over twenty sources: official financial filings via XBRL feeds, sell-side analyst estimates, alternative data providers (think credit card transaction aggregates, web scraping of product reviews, and satellite imagery analysis), and macroeconomic databases from central banks. Each source comes with its own quirks. For instance, Chinese company financials reported under IFRS can differ materially from those reported under US GAAP, and we've had to build normalization layers that account for these discrepancies. One particularly sticky issue was revenue recognition in the SaaS sector—different firms recognize multi-year contracts differently, and naive aggregation would create false signals.
**Quality control is an ongoing war**, not a one-time battle. We employ a tiered validation system. First, automated sanity checks: does total assets equal liabilities plus equity? Is revenue growth positive but not exceeding 1000% year-over-year? Second, cross-source reconciliation: if Company X's management says they have 10 million users in their earnings call, but a third-party data provider shows only 8 million active users, we flag the discrepancy for human review. Third, we maintain a "data pedigree" database—essentially a blockchain-like ledger (without the crypto hype) that tracks every transformation applied to raw data, allowing full auditability. This matters because when a model makes a weird call, we need to trace it back to potential data contamination.
Let me give you a real case. Early last year, our system started generating inexplicably bullish signals for a mid-cap logistics company in Southeast Asia. The fundamentals looked stellar—growing margins, declining debt, expanding market share. But something felt off to our senior analyst, Mei-Lin. She dug into the data provenance and discovered that the alternative data feed we used for tracking truck utilization had double-counted certain routes due to a software update bug. The actual utilization was 15% lower than reported. Without the quality control layer, we would have acted on a phantom signal. This experience reinforced our commitment to data integrity as a non-negotiable prerequisite. We now track over 200 quality metrics per data source, and any source that falls below a 95% reliability score over a rolling quarter triggers an automatic review.
Moreover, we've developed a feedback loop where the system itself learns to identify suspicious data patterns. For example, if a company's reported accounts receivable suddenly drops by 40% while days sales outstanding remains constant, our anomaly detection module flags it. This isn't foolproof—we've had false positives where legitimate restructuring explained the change—but it forces human oversight at exactly the right moments. The cost of building this infrastructure is substantial, pushing into seven figures annually for a mid-sized firm like ours. But compared to the cost of a single bad trade based on flawed data, it's cheap insurance.
---Factor Construction and Selection
Once data is clean, the next challenge is turning raw numbers into meaningful factors. In quantitative finance, a factor is a measurable attribute that explains differences in asset returns. Classic examples include value factors (price-to-book ratio), momentum factors (past six-month returns), and quality factors (return on equity). But our system goes beyond these textbook definitions. We construct *contextual factors* that adapt to industry, market regime, and company lifecycle stage.
Here's the thing—most off-the-shelf factor libraries are built on US equity data from the 1960s onwards. They work reasonably well in developed markets but break down in emerging markets or for small-cap stocks. At DONGZHOU LIMITED, we operate across multiple geographies, including China, India, and Brazil. A value factor that works in the S&P 500 often fails in the Shanghai Composite because of different market structures—state-owned enterprises don't behave like private firms, and retail investor sentiment dominates institutional flows. So we had to build *bespoke factors*.
**One factor I'm particularly proud of** is what we internally call the "Earnings Quality Deception Index." Traditional metrics like accruals ratio capture some manipulation risk, but we found that combining textual analysis of earnings call transcripts with financial statement anomalies creates a more powerful signal. The NLP model scans for certain linguistic patterns: excessive use of passive voice, avoidance of specific numbers, overuse of words like "transformation" or "synergy" without concrete details. When this textual score aligns with financial red flags—like growing divergence between operating cash flow and net income—the combined factor predicts negative earnings surprises with 65% accuracy, compared to 52% for the standalone financial metrics.
Factor selection itself is a delicate art. We maintain a library of over 300 candidate factors, but using them all would overfit and kill performance. Our selection process involves three stages. First, we run each factor through a battery of statistical tests: t-statistics, Fama-MacBeth regressions, information coefficients, and turnover rates. Factors must demonstrate statistical significance at the 5% level over multiple market cycles. Second, we assess *economic plausibility*—a factor might be statistically significant but make no logical sense. For instance, we once found that companies with CFOs named "Michael" outperformed. It was statistically significant in our sample due to data mining, but we rightly rejected it. Third, we test for robustness across different time periods, geographies, and market conditions. A factor that works in bull markets but crashes in bear markets might still be useful, but we need to know its conditional properties.
Our research draws heavily on the work of academics like Eugene Fama and Kenneth French, but we also incorporate more recent findings from behavioral finance. Richard Thaler's work on limited attention informs our "neglected stock" factor, which identifies companies with low analyst coverage relative to their market cap. The logic is simple: underfollowed stocks have a higher probability of being mispriced because fewer smart eyes are watching. Empirically, this factor generates about 2-3% annualized alpha in our universe, though with higher volatility. We also use insights from Andrew Lo's "Adaptive Markets Hypothesis" to dynamically adjust factor weights based on market regime—value factors get higher weights in contraction phases, while momentum factors dominate in expansions.
---Model Architecture and Ensemble Methods
Now we arrive at the engine room. How do we actually combine these factors into a predictive model? I could give you the textbook answer—machine learning, gradient boosting, neural networks—but the reality is more nuanced. At DONGZHOU LIMITED, we've settled on an **ensemble architecture** that blends multiple modeling approaches, each with its strengths and weaknesses. The goal is not to build the single best model, but to create a robust system that degrades gracefully when individual components fail.
Our system has three primary model layers. The first is a **fundamental regression layer**, built on traditional econometric techniques like Fama-MacBeth cross-sectional regressions and panel data models. These models are transparent, interpretable, and statistically rigorous. We can trace exactly which factor contributed how much to each prediction. The downside is that they assume linear relationships and fixed coefficients, which is rarely true in financial markets. During the 2020 COVID crash, for instance, the relationship between leverage and returns flipped from negative to positive—highly leveraged companies with strong cash positions outperformed, while the traditional model predicted the opposite. The linear model failed precisely when we needed it most.
The second layer uses **gradient boosted trees** (specifically LightGBM and XGBoost) to capture non-linear interactions. These models excel at learning complex patterns like: "Companies with moderate debt, high insider ownership, and positive earnings revisions outperform only when interest rates are below 3%." The trees can automatically discover these interactions without manual specification. However, they are prone to overfitting, especially in the small-sample regimes typical of emerging markets. We combat this through aggressive regularization, early stopping, and out-of-time validation—never out-of-sample, always out-of-time, because financial data has temporal dependencies that random splits miss.
The third and most experimental layer is a **transformer-based neural network** that processes sequential financial data. Think of it as a small language model, but instead of words, it ingests time series of financial ratios, macroeconomic indicators, and textual sentiment scores. We adapted the architecture from Google's BERT model, training it on eight years of quarterly data across 5,000 global stocks. The attention mechanism allows the model to weigh the importance of different time steps—for example, recognizing that a sudden spike in accounts receivable three quarters ago is more predictive of current earnings quality than the most recent quarter's data. The neural layer adds about 3-5% predictive improvement over the gradient boosting alone, but at a computational cost that makes it impractical for real-time trading. We use it primarily for weekly portfolio rebalancing decisions.
**The ensemble combines these layers through a meta-learner**—a simple logistic regression that learns the optimal weight for each model's prediction based on recent performance. If the neural network has been outperforming in volatile markets, its weight increases. If the linear model is misfiring, its contribution shrinks. This dynamic weighting is crucial. We also maintain "model health dashboards" that monitor prediction errors, feature importance drift, and model confidence levels. When a model's confidence drops below a threshold—say, 60%—the system automatically reduces its influence and flags the episode for human review. This isn't perfect; we've had instances where all three models agreed on a wrong prediction because a structural break (like a regulatory change in China's tech sector) hadn't been captured in the training data. But the ensemble approach at least prevents catastrophic single-model failures.
---Risk Management Integration
No quantitative system is complete without rigorous risk management. But I don't mean the standard Value-at-Risk (VaR) calculations that banks use for regulatory compliance. I mean *dynamic, forward-looking risk measurement* that accounts for factor correlations, tail dependencies, and regime changes. At DONGZHOU LIMITED, we've built what we call a **"Factor-Risk Decomposition Engine"** that sits alongside the predictive model.
The core idea is simple but powerful: every portfolio position is decomposed into its exposure to fundamental factors (value, momentum, quality, size, etc.) and the residual "idiosyncratic" risk. We then measure not just the volatility of each factor, but the *stability of factor correlations*. During calm markets, value and quality factors might have a correlation of -0.2. But during crises, correlation can spike to 0.6 as everything collapses together. Standard risk models that assume constant correlations severely underestimate tail risk. Our system uses a regime-switching model that estimates probabilities of being in a "normal," "stressed," or "crisis" regime, and adjusts risk forecasts accordingly.
**One practical application** is our position sizing algorithm. Most quantitative funds use a fixed risk budget—say, 2% of capital per position. We instead use *dynamic risk parity* where position size is inversely proportional to the factor risk contribution. If a stock has high exposure to a factor that is currently volatile (like the "momentum factor" during a reversal), its position size is automatically reduced. Conversely, if a stock offers unique exposure to a stable factor, we can overweight it. This approach reduced our maximum drawdown during the 2022 bond market rout from -18% to -11%, simply because we dynamically scaled down positions with high exposure to rising interest rates.
We also incorporate what I call "narrative stress testing." Traditional stress tests apply historical shocks—like a 20% market decline or a 2% interest rate hike. But these are backward-looking. Instead, we generate *plausible future narratives* using large language models, then stress the portfolio against those scenarios. For instance, we recently tested a scenario where "Supply chain disruptions worsen due to Taiwan Strait tensions, causing semiconductor companies to lose 30% of revenue for two quarters." Our system had to estimate how each factor would behave—likely, the "quality factor" would strengthen as investors flee to profitable companies, while the "small-cap factor" would weaken. This allowed us to preemptively reduce exposure to small-cap semiconductor suppliers. The narrative approach is speculative, but it forces us to think about *unknown knowns* rather than just known risks.
**Academic research supports our approach.** The work of Campbell Harvey and Yan Liu on "skeptical risk management" shows that quant funds that actively manage factor exposures outperform those that rely on static risk models by about 2% annually over 20-year periods. We've also drawn lessons from the collapse of Long-Term Capital Management in 1998—their failure wasn't just about leverage, but about assuming that historical correlations would hold during stress. Our regime-switching model explicitly addresses this by constantly testing the stability of correlation structures.
---Execution and Implementation
Having a great model is useless if you can't trade on it. The gap between a theoretical portfolio and actual returns is filled with **implementation shortfall**—the costs and slippage that eat predicted alpha. At DONGZHOU LIMITED, we've developed a dedicated execution module that bridges this gap. It's not glamorous work, but it's where real profits are made or lost.
The first challenge is **timing**. Our model might identify a stock that should be bought based on quarterly data released yesterday. But by the time we analyze, decide, and execute, the market may have already adjusted. We combat this through a "pre-market scoring" system that runs our models on overnight data releases, generating trade signals before the Asian market opens. For US-listed stocks, we score after close for the next day's execution. The latency from signal generation to execution is under 15 seconds for liquid stocks, and under 2 minutes for less liquid names. This speed advantage captures about 20-30% of the alpha that would otherwise be lost to time decay.
The second challenge is **market impact**. Our portfolio holds positions in mid-cap stocks where a large order can move prices against us. We use a "volume-weighted average price" (VWAP) execution algorithm that slices big orders into smaller chunks, executed over several hours or even days. The algorithm is adaptive—if volatility spikes, it slows down execution; if liquidity is plentiful, it speeds up. We've tuned this algorithm using reinforcement learning, training it on historical order book data. The model learned, for example, that executing during the first 30 minutes of the US market open is expensive due to high volatility, and that the last hour often offers better liquidity for exiting positions. These micro-optimizations save us about 0.5% in execution costs annually—not huge on its own, but compounded over years and across 50-100 trades per month, it adds up.
**A case from our experience**: In mid-2023, our model generated a strong buy signal for a Korean battery manufacturer as electric vehicle adoption accelerated. But the stock was thinly traded, with daily volume of only $5 million. Our standard execution algorithm would have taken a week to build a full position, risking price movement against us. Instead, we used a "dark pool" access provided by our prime broker to execute large blocks anonymously, combined with a patient approach that used only 10% of daily volume each day. The position took eight days to build, but the average execution price was within 0.3% of our target. An impatient execution would have cost us 2-3% in impact costs. This is the kind of granularity that separates professional execution from amateur guesswork.
We also monitor **short selling constraints**. Some of our models generate short signals, but in many emerging markets, shorting is expensive or impossible due to securities lending constraints. We maintain a database of short availability across our universe, updated in real-time from our prime brokers. If a stock isn't borrowable, the system automatically converts the short signal into a "relative underweight" in the portfolio, reducing allocation rather than going negative. This pragmatic adjustment ensures we're not chasing theoretical trades that can't be executed in practice.
---Backtesting and Validation
I've seen more phony backtests than I care to count. People take a dataset, fit a model, report Sharpe ratios of 3.0, and call it a day. Then the model blows up in live trading. **Proper backtesting is a rigorous science**, and at DONGZHOU LIMITED, we've institutionalized a culture of skeptical validation.
Our first rule is **out-of-time validation**. We never use random train-test splits because financial data is temporally dependent. Instead, we train on data from 2010-2019, validate on 2020-2021, and test on 2022-2023. This ensures the model must predict periods it has never seen. One shocking finding: many promising factors that worked in the 2010s completely broke down during the 2022 inflation shock. Our "low volatility" factor, which historically predicted strong performance, turned negative as rising rates crushed stable but leveraged companies. Had we only done random splits, this failure would have been invisible. Now, we always require that a strategy passes at least two distinct time periods—one bull, one bear—before we consider it live-ready.
Second, we simulate **realistic trading constraints**. The backtester includes transaction costs (commission, spread, market impact), slippage, short availability, position limits, and cash drag. We also model the time delay between signal generation and execution—if a signal is based on earnings reported after market close, the backtester should execute at the next day's open, not at the close of the same day. These details matter enormously. A strategy that shows 15% annualized return in a naive backtest might drop to 8% after realistic costs, and to 4% after accounting for capacity constraints (i.e., the strategy works for $10 million but not $100 million). We actually run our backtester at three different capital levels to understand scalability.
Third, we conduct **Monte Carlo simulations with bootstrapped residuals**. Instead of assuming returns are normally distributed, we sample from the actual empirical distribution of our model's forecast errors. This generates thousands of possible paths for portfolio evolution, allowing us to estimate not just expected return, but the distribution of outcomes. This is where I learned humility. Our flagship value strategy had an expected Sharpe of 1.2, but the Monte Carlo simulation showed a 15% probability of losing money over any two-year period. That 15% tail risk was invisible in the point estimate. We now use these simulations to set stop-loss triggers and maximum drawdown limits.
Research from Marcos López de Prado on "Deflated Sharpe Ratios" has been particularly influential in our lab. He shows that the number of trials (how many strategies you test) dramatically inflates apparent performance if you don't correct for multiple testing. We maintain a log of every strategy we've ever tested—over 1,200 at last count—and adjust our significance thresholds using the False Discovery Rate (FDR) method. This is painful because it means many "promising" strategies get rejected. But it saves us from the overfitting that has destroyed many quant funds.
.png)
Finally, we run **paper trading for minimum three months** before deploying any model with real capital. The paper trading environment is identical to live—same data feeds, same execution delays, same risk controls. The only difference is that profit and loss is simulated. In 2022, we had a model that looked great in backtests and passed all validation. In paper trading, it generated 25% of the expected return. Investigation revealed that our backtester had accidentally used a forward-looking bias in the earnings data—it was applying revised earnings figures that wouldn't have been known at the trade time. This data leakage is embarrassingly common in our industry. The paper trading period caught it before real money was at stake.
---Human Oversight and Behavioral Safeguards
Here's a confession: I used to believe that a fully automated system was the holy grail. Remove humans, remove bias, right? Wrong. **Humans are terrible at fighting their own biases, but machines are terrible at adapting to novel situations.** The optimal architecture is a *human-in-the-loop system* where the machine handles routine decisions and the human steps in for exceptions, validation, and strategic overrides.
At DONGZHOU LIMITED, our daily workflow looks like this. The system generates a list of recommended trades each morning. A senior analyst reviews the top ten signals, reading the underlying fundamental reasoning—why is this stock cheap? Is there a catalyst? Are there any known risks the model might have missed? The analyst can override any trade, but the override must be documented with a reason. We track these overrides and analyze them quarterly. Interestingly, we've found that analysts are good at catching *false positives* (trades that look good but have hidden risks) but tend to miss *false negatives* (missed opportunities). The machine outperforms humans in breadth, humans outperform in depth. Together, they form a powerful team.
**Behavioral safeguards are embedded in the system design.** For instance, we implemented a "confirmation delay" for large trades—any single order exceeding 5% of portfolio value requires a 30-minute waiting period and a second confirmation. This prevents impulsive reactions to news or market movements. We also have a "circuit breaker" that automatically halts trading if daily losses exceed 2% of portfolio value. The system then enters a "review mode" where no further trades are allowed until a risk committee meets and documents the cause of losses. This forced pause has saved us from multiple tailspin scenarios. In March 2023, during the Silicon Valley Bank crisis, our circuit breaker triggered after a 2.5% loss in a single day. The review revealed that our liquidity factor had flipped unexpectedly, and we needed to adjust our risk model. Without the circuit breaker, we might have doubled down and suffered deeper losses.
We also conduct **quarterly "post-mortems"** of every losing trade. This isn't about blame; it's about learning. We categorize each loss into one of five buckets: model error (incorrect prediction), data error (bad input), execution error (slippage worse than expected), external shock (unforeseeable event), or psychological error (analyst override that was wrong). Over two years, we found that 40% of losses came from external shocks, 30% from model error, 15% from data error, 10% from execution, and only 5% from psychological error. This data helped us focus improvement efforts—we invested in better tail-risk modeling and expanded our data validation team. The human element was already well-controlled.
I draw inspiration from Daniel Kahneman's work on cognitive biases. His concept of "System 1 vs System 2 thinking" maps directly to our design. The quantitative model handles the fast, automatic, high-frequency decisions (System 1). The human analysts engage in slow, deliberate, analytical reasoning for the important exceptions (System 2). The key is knowing *when* to switch between systems. We've trained our analysts to recognize "red flag" patterns—a sudden spike in unexplained volatility, a concentrated bet in correlated positions—that trigger deeper human scrutiny. This hybrid approach isn't perfect. We still miss things. But it's substantially better than either pure automation or pure discretion.
--- ## Convergence and the Path Forward As I sit here at my desk in DONGZHOU LIMITED's Shanghai office, watching the real-time performance dashboards flicker with our portfolio's P&L, I'm struck by how far we've come. The Fundamental Quantitative Analysis System we've built is not a single product, but a living, evolving organism. It's part data pipe, part statistical engine, part risk sentinel, and part human augmentation tool. Each component—from data quality to factor construction to model architecture to risk management to execution to validation to oversight—interacts with the others in complex, sometimes unpredictable ways. The central insight that has guided our work is this: **there is no free lunch in quantitative finance**. Every improvement in predictive power comes with a cost—whether in data procurement, computational resources, or complexity risk. The systems that survive are not necessarily the ones with the highest Sharpe ratios in backtests, but the ones with the most *robust* architectures—those that can withstand data failures, regime changes, and human errors. We've built redundancy into every layer. We've embraced the uncomfortable truth that we will never know the "true" model, and so we build ensembles that are resilient to model misspecification. Looking forward, I see three frontier areas where our system will evolve. First, **integration of causal inference methods**—moving beyond correlation to identify causal drivers of returns. This is hard in finance because controlled experiments are impossible, but techniques like instrumental variables and difference-in-differences show promise. Second, **explainable AI** that not only predicts but *reasons* about why a stock is mispriced. Our current models are black boxes; investors (and regulators) increasingly demand transparency. Third, **real-time adaptation** using online learning algorithms that update models as new data streams in, without requiring full retraining. The world changes too fast for quarterly model updates. But the core philosophy remains unchanged: fundamentals matter. The price of a stock eventually converges to the present value of its future cash flows. Everything else—sentiment, momentum, technical patterns—is noise around that fundamental truth. Our system doesn't try to predict the noise; it tries to estimate the signal with ever greater precision. And in that pursuit, we find not just profits, but genuine understanding of how businesses and markets work. --- ## DONGZHOU LIMITED's Insights on the Fundamental Quantitative Analysis System At DONGZHOU LIMITED, our journey building this system has taught us that **the competitive advantage in quantitative finance lies not in secret formulas or proprietary data, but in operational excellence and intellectual honesty**. Many firms claim to have "quantitative fundamental" systems, but few have the discipline to maintain the data quality pipelines, the validation rigor, and the human oversight that make such systems actually work in practice. We've learned that the biggest risks are not market risks, but operational risks—data bugs, model overfitting, execution glitches, and cognitive biases creeping into supposedly objective systems. Our approach has been to institutionalize skepticism: every output must be challengeable, every assumption must be testable, and every failure must be learnable. We do not claim to have found the "truth" about market efficiency. Rather, we've built a systematic process for asking better questions, and that process—not any single model—is our core asset. As the global financial ecosystem becomes more interconnected and data-rich, we believe that the firms which master the *integration* of fundamental understanding with quantitative rigor will outperform those who specialize in either alone. This is not just a technological challenge; it is a cultural one. We are building a culture that values both mathematical precision and fundamental business judgment, and we invite our partners to join us in this exploration. ---