Quantifying Narrative Similarity with VecEvent Correlation

December 2, 2025

Correlation measured using trailing returns is volatile and backward looking

Many investors measure diversification via historic return correlation, which is itself correlated to volatility. When incorporated into portfolio risk metrics such as VaR, it can encourage selling amidst panics and buying amidst complacent booms.

VecEvent based correlation is more stable and very much “story” driven

VecEvents are news events and themes that influenced a ticker’s price. VecViz links a VecEvent to a ticker’s Vector Set channel when one or more of the dates of the tops and bottoms anchoring the channel overlap with the VecEvent’s period of influence.

VecEvent based correlation is at its core qualitative and text based, with NLP bridging it to the numerical realm. It goes beyond measuring industry and issuer concentration (as powerful as those considerations can be¹). It measures concentration of theme and event related sensitivity. In an age of disruption that can cut across industries, be it policy (ex: tarriffs) or technology (ex: ai) driven, its theme and event orientation can provide important perspective on diversification.

NLP in Finance: A Decades-Long Pursuit Entailing Serious Data Wrangling

While quants have long scraped news for sentiment (bullish/bearish), few have used it to map structural similarity between assets. Modern Natural Language Processing (NLP) techniques move beyond simple common word counts to extract deep, contextual meaning. This capability can turn vast amounts of unstructured text into financial signals, offering a fundamentally different way to model asset relationships.

However, the task of sourcing, cleaning, labelling, and processing a vast, unfiltered corpus of global news, filings, and articles remains formidable.

The VecViz Approach: LLM Curation of a Vast Text Corpus Meets Transparent NLP

The VecViz methodology bypasses the heavy text data procurement and processing lift by sourcing its VecEvevents through an LLM prompt, which delivers them with accompanying ticker and influence start and end date fields². It then applies a mature, somewhat transparent NLP pipeline to distill common theme “benchmarks” across each ticker’s VecEvents and score each ticker’s exposure to each theme. Correlation is calculated on the basis of exposure across all the benchmarks.

VecViz’s VecEvent Correlation Calculation Pipeline

1. Turn Event Text into Numbers (TF-IDF): The process starts by building a “dictionary” of all the unique, important words across all VecEvents of all companies under consideration. It then scores each VecEvent’s exposure to each word. We do this with sklearn’s Term Frequency-Inverse Document Frequency (TF-IDF) vectorizer function. Words that are common in one event but rare overall (like “acquisition”) get a high score, making them stand out.
2. Find Common Theme Benchmarks Across VecEvents: After all events are converted to numerical fingerprints, this step automatically sorts them into groups based on their similarities using K-Means Clustering. The center of each group (a “centroid”) becomes the numerical definition of a common theme or “benchmark,” such as “Legal Issues” or “New Product Launches.”
3. Create a Single Profile for Each Company: To move from the event level to the company level, the process creates one “summary profile” for each firm. It does this by looking at all the key word exposure scores for a company’s VecEvents and taking the peak score for each word. This ensures that rare but critical events (like a major drug approval) define the company’s profile.
4. Score How Much Each Company Fits Each Theme: With a single profile for each company, this step measures how “aligned” each company is to each of the themes found in Step 2. It uses cosine similarity to generate a “benchmark exposure profile” (like a signature) for every firm, showing which themes (e.g., 70% “Acquisition” theme, 20% “Legal” theme) it’s most associated with.
5. Calculate How Similarly Companies Behave (Pearson Correlation³): Using these new “benchmark exposure” signatures, the process calculates a correlation score between any two companies.

This process is arguably also akin to risk factor modeling, with the VecEvent Benchmarks playing the role of the risk factor return indices.

Optimal calibration may reside at 1.5 benchmarks per ticker, given 10 VecEvents/ticker

The number of VecEvent benchmarks utilized is a key driver of the results obtained. We ran the VecEvent correlation on the training data for the 138 tickers studied for all VecEvent Benchmark counts between 2 and 200 to see which benchmark count yielded the lowest MAE to corresponding Pearson correlation of trailing 252d returns as of 4/30/2024. The results are charted below. Diminishing returns really kick in at just under 1 benchmark per ticker (note we had an average of ~10 VecEvents per ticker), but the optimal quantity was reached at 192, about 1.5 benchmarks per ticker.

VecEvent correlation skewed a bit lower than T252d return based Pearson correlation

In sample results for VecEvent correlation using the 192 benchmarks (green) are compared to corresponding Pearson correlation (blue) in the chart below. Obviously, the fit to Pearson is fairly strong, though it does skew lower. Also pictured is VecViz’s fingerprint correlation method, which you can read more about here.

VecEvent based correlation outperformed alternative correlation measures in recent study

We did a study of VecViz analytics, including the VecEvent based correlation metric discussed above, as inputs to portfolio mean-variance optimization (MVO)⁴. In this study we compared the VecViz inputs to each other and to simple trailing 252d return based alternatives. You can find the full study, entitled “VecViz Analytics Performance as MVO Portfolio Optimization Inputs” here.

The study ranks portfolios based on a metric we call “SummaryZ”, which contemplates standardized scores for Annual Average Return, Max Drawdown, Sharpe Ratio, Calmar Ratio⁵, Multi-Factor Alpha⁶, and Kupiec P-Value⁷.

The table below, excerpted from the report, indicates the results across a grid search of constraints and rebalance frequencies. The VecEvent based correlation process described here is denoted as “Correl (VE)_VV”. It outperforms both rolling 365 day Pearson correlation, denoted as “Correl_T252d”, and VecViz analytic “fingerprint” based correlation , denoted as “Correl(FP)_VV”.

Key to Input Variable abbreviations:

• Vol_VV = VecViz’s 99D_Ret (i.e. VecViz 99% VaR)
• Correl(VE)_VV = VecEvent based correlation
• Ret_T252d= average price return over the prior 252 days
• Correl(FP)_VV = VecViz analytic metric “fingerprint” based correlation
• Ret_VV = VecViz’s VaR and OaR breakage regime based expected return metric
• CorrelT252d= correlation of price returns over the prior 252 days
• Vol_T252d = standard deviation of price returns over the prior 252 days

That said, look ahead bias cannot be ruled out when using retrospective LLM output

We are confident that this data was out of sample with regard to the Correl (FP)_VV and Correl_T252d portfolios. However, it is impossible to be entirely confident that it was out of sample with regard to the VecEvent Correlation based portfolios (denoted in the table as Correl (VE)_VV) given that the prompts generating the VecEvents occurred during the test period. We attempted to mitigate look ahed bias by excluding VecEvents with influence start dates that began during the test period but nevertheless, we can never know that the same set of VecEvents would have been generated had the prompt occurred prior to the start of the period tested.

Conclusion

Even accounting for some potential LLM look ahead bias, the performance gap between VecEvent based correlation and the alternatives suggests that it should be at least “in the mix” of the metrics used to estimate forward correlation. “Price is truth”, as the saying goes, but the story behind the price chart seems like it should have greater intrinsic and enduring meaning than the chart itself. If that is so then considering the story behind tickers when assessing diversification can help build more robust investment portfolios.

My thanks to Fathmat Bakayoko for her help with the development and calibration of VecEvent correlation.

Notes:

1. The CLO Market’s “Diversity Score” was an early text-driven diversification metric. For Moody’s, the ratings of CLO debt is based in part upon a “Diversity Score” of the loan collateral, which considers industry and issuer concentration. Other ratings agencies have similar metrics or embedded considerations as key parts of their ratings process.The default rate of CLO’s has been much lower than that of comparably rated US corporate credit over the 1997-2023 period. To what extent this strong default rate performance is attributable to the Diversity Score (et al) vs. other CLO rating considerations is unclear. However, in the context of still widespread usage of trailing return based correlation for expected portfolio volatility assessment it arguably deserves broader consideration than it has received to date, if only as a springboard for a better covariance-friendly measure. ↩︎
2. Firms that have fundamental analysts with deep knowledge of the tickers they cover can choose to source their VecEvents from their analysts, and may find even greater alpha in doing so. ↩︎
3. A standard measure (-1 to +1) of linear relationships. At its core is summing the product of differences from the mean for two sets of data that are linked in some way (ex: they relate to the same benchmark). Essentially, it quantifies how often two variables are on the “same side” of their respective averages simultaneously, giving greater weighting to larger deviations. ↩︎
4. Mean-Variance Optimization (MVO) is a quantitative tool used to construct portfolios. It weighs three factors: the expected return of an asset (the Mean), the risk of that asset (the Variance) and how the asset relates to other assets under consideration (correlation). MVO identifies the portfolio with the maximum expected return for a given level of risk and other user defined constraints (ex: max ticker weighting) ↩︎
5. Calmar Ratio = Average Annualized Return / Max Drawdown ↩︎
6. Multi-Factor Alpha is determined via mutliple regression of MVO generated portfolio returns on the returns of MTUM, VLUE and SPY (the iShares MSCI USA Momentum Factor ETF, the iShares MSCI USA Value Factor ETF, and the SPDR S&P 500 Trust ETF, respectively) ↩︎
7. “Kupiec P-Value” = Kupiec Test Statistic P-Value, which here reflects the probability that the portfolio’s 99% VaR, as implied by its volatility constraint, (assuming normality, and independent, identically distributed daily returns) was well specified. ↩︎