VecViz V-Score (VS) Performance Report Summary: 1/30/2025

VecViz’s performance reports analyze the behavior and effectiveness of our analytics. This post focuses on the ‘VecViz V-Score (VS) Performance Report’ as of February 1, 2025. It summarizes key findings and explains how to use the report to investigate the performance of specific tickers.. You can find this report on the “Reports” page of the VecViz website, along with updated reports over time.

The V-Score is a ranking of expected forward price returns that ranges from -12 (bearish) to +12 (bullish).The V-Score is calculated using machine learning techniques applied to Vector Model inputs and outputs across six time horizons (from 1 day to 1 year)¹. Each horizon is assigned a quintile ranking score ranging from -2 and +2. The final V-Score is the sum of these horizon scores, thus ranging from -12 to +12. The full frequency distribution of out of sample V-Score observations is depicted below.

V-Scores calculated as of the prior day’s close can be found on the “Dashboards” page of the VecViz website. Detailed descriptions of each of the features contributing to the V-Score can also be found there, by hovering your mouse over the labels in the spider chart, an example of which is pictured above.

Performance Metrics Utilized in the V-Score Report

The primary metric used in the report is average price return. A secondary metric used in the report is Information Ratio (IR), which we calculate by dividing the average return by the standard deviation of returns². The information ratio allows comparison on the basis of average return adjusted for return volatility.

The report allows for easy comparison between V-Score interval groupings of ticker-model dates (TMD’s). Price returns and IR’s for the average ticker for which VecViz generates V-Scores (“AvgTicker_VV”), and to well known ETF’s such as “SPY” and “QQQ” are also provided, for comparison purposes.

Ticker Groupings Utilized in the V-Score Report

Some V-Score interval TMD groupings are broad, such as “PosVS” and “NegVS”, which represent all positive and negative VS, respectively. Others are much, much more narrow, such as VS>9 and VS <-9, representing TMD’s with VS greater than 9 and less than -9, respectively. These groupings represent only 1.2% and 0.4% of all TMD’s, respectively. Even VS>6 and VS<-9 are rather thin, representing only 5.7% and 2.1% of TMD’s respectively. Note that the thinner diversification of such groupings makes their performance more vulnerable to the influence of outliers.

To illustrate the ability to utilize the V-Score in conjunction with other metrics, we also include V-Score groupings that take into consideration VaR as well: “PosVaRAdjVS (Bull)” and “NegVaRAdjVS (Bear)”.

• The PosVaRAdjVS grouping includes ticker-model dates for which the ticker has a positive V-Score that is greater than expected given its 95% VaR level.
  • The rationale here is simple – tickers with the most bullish (V-Score based) outlooks relative to their downside.
• The NegVaRAdjVS grouping includes ticker-model dates for which the ticker has a negative V-Score that is less negative than expected given their 95% VaR level.
  • The rationale here is more subtle – tickers with at least a modestly bearish (V-Score based) outlook, for a more deeply bearish outlook may be justified given their downside.

Finally, to illustrate the magnitude and stability of return differentials between “bullish” and “bearish” V-Scores we also include in the tables featured below two “Long-Short”³ groupings: “PosNeg_Diff” and “VaRAdjPosNeg_Diff”.

• “PosNeg_Diff” reflects the performance differential between “PosVS (Bull)⁴” and “NegVS (Bear)⁵ groupings
• “VaRAdjPosNeg_Diff” reflects the performance differential between the “PosVaRAdjVS (Bull)” and “NegVaRAdjVS (Bear)”, groupings.

V-Score Performance Summary Table Example

In the tables below we display the performance metrics for the V-Score based groupings, as discussed above.

The table below includes every ticker in VecViz’s coverage universe for every model date (i.e., every US equity trading day) in the entire 3 year out of sample period ending 1/30/25. The columns represent each of the six investment time horizons considered (denominated in trading days, where 21d ~= 1 month, 252d ~= 1 year). 

See p.20-p27 of the report for additional detail on these figures and how they vary if certain ticker groups are excluded. To see how similar detail for the trailing 1 year, 3 month, and 1 month time lookback windows see p.28-48 of the report. To see the variability of the price returns for each bullish V-Score group and its bearish counterpart charted by model date, see pages 49-78 of the report.

V-Score “Report Card”

To evaluate the robustness of the V-Score’s behavior we ran the analysis above for a range of lookback window and ticker universe variations and evaluated the results on the basis of their consistency with expectations. We list those criteria in the sections that follow and then present the resulting “Report Card”, that indicates the % of the weighted expectations that were met.

Each column of the table is a time horizon and each row is a ticker grouping described by the tickers excluded and lookback window (from 1/30/2025) combination. For example, the Average Price Return and IR heatmaps in the prior section represent all tickers for the entire out of sample performance history. The evaluation of the results presented in those heatmaps appears in the “None_AllD” row of the Report Card below (because tickers excluded are “None” and the lookback window includes “All” model dates).

Average Price Return (+2 if met):

• Positive V-Scores > Avg Ticker > NegV-Scores (i.e. PosVS (‘Bull’)>AvgTicker_VV> NegVS (‘Bear’), to use the nomenclature of the heatmaps above)
• Positive VaR Adjusted V-Scores >Average Ticker > Negative VaR Adjusted V-Scores
• Positive V-Score Rank Order corresponds to Price Returns
• Negative V-Score Rank Order corresponds to Price Returns
• The differential between Positive and Negative V-Scores is greater on a VaR adjusted basis than on an unadjusted basis.

Information Ratio (+1 if met):

• Positive VaR Adjusted V-Scores > All Positve V-Scores
• Negative VaR Adjusted V-Scores < All Negative V-Scores
• Positive VaR Adjusted V-Scores> Average Ticker
• Negative VaR Adjusted V-Scores < Avg Ticker
• Positive VaR Adjusted V-Scores > SPY

Observations:

• HORIZONS: Performance has degraded over the last 365 days relative to the “Alld” (prior 3 year) period, most significantly for the 63d and 126d horizons. There are only 4 unique 63 trading day horizons in a year, and only 2 unique 121 trading day horizons. Modelling them is by nature perilous. Degredation for other horizons over the last 365 days was much less severe.
  • Performance degraded a bit further over the last 90 days relative to the last 365 for the 1d and 21d horizons, though the 10d horizon held in well.
  • Even with the recent dip in the last 30d’s (noted on the Report Card), the 10d horizon is likely the most suitable time horizon for the V-Score, followed closely by the 21d horizon.
• V-Score Groupings: The groupings “VS <-6” and “VS<-9” have much stronger average returns than expected. This is the most glaring weakness in the V-Score’s performance. Fortunately very few tickers fall into these groupings, as discussed previously.
  • Much (though not all) of this surprisingly strong performance was driven by the “Crypto/Mem” ticker grouping. A bullish back drop has likely exacerbated it, as the dynamic was even more pronounced in 2023 and 2024 than it was in 2022.
• The VaR Adjusted Positive V-Score grouping (PosVSVaRAdj (Bull) is the most interesting from the bullish side, outperforming the SPY for the 10d horizon in several of the ticker grouping and lookback window combinations considered.
• The VaR Adjusted Negative V-Score grouping (NegVSVaRAdj (Bear) is the most compelling from the bearish perspective.
• Ticker Groupings: Semi’s contributed the most to performance, Small Caps detracted the most, in aggregate.
  • As discussed previously, the “Crypto/Mem” grouping drove a large part of the unexpectedly strong average price return performance of the VS<-6 and VS<-9 V-Score groupings.

Appendix 1: Ticker Level V-Score Performance Report Inquiry Example

When we press “Ctrl-F” on the V-Score report pdf file and enter “NVDA”, for example we learn that it appears 77 times in the document. The first appearance is in the list of tickers covered, which is provided in the Introduction.

• The next 59 occurences are in “Top 30 Ticker” lists, which appear between p.60 and 140. Tickers listed on the left column were expected to be bullish – these tickers performed as expected. NVDA fortunately is in the left column within this section of the report 59 out of 59 times.
• The “Bottom 30 Ticker” lists start on p.141, and NVDA appears 17 times in this section. Tickers that are listed on the right side of the page in this section performed as expected – the expectation for them was bearish. NVDA appears in the right side column 6 of the 17 times it appears in this section.

Thus, when the model was bearish on NVDA and its performance was sufficently volatile to be in the Top or Bottom 30 tickers for one of the lookback windows and one of the 6 horizons studied, it was 100% accurate. When it was bullish on NVDA and those same volatility conditions were met it was correct 59 out of 70 times, or 84% accurate.

1. The training period for the V-Score includes 250 model dates randomly selected from June 2005 to January 2021. Tickers included in the V-Score training data include nearly all the tickers presented in this report. There has been no “roll forward” of the training data since then. See the full V-Score performance report for more detail. ↩︎
2. Note that standard deviation of returns is understated for horizons greater than 1d, as it is calculated upon overlapping periods. However, since this dynamic applies uniformly across all tickers considered standard deviation can still be considered on a comparative basis within the confines of this report. ↩︎
3. We include no cost of borrow or repo earnings or margin cost in the returns for these Long-Short groupings, just the simple difference between the average metric for the bullish grouping and the corresponding bearish grouping. ↩︎
4. Includes all TMD’s with a positive V-Score. ↩︎
5. Includes all TMD’s with a negative V-Score. ↩︎