This article is also my quant research workbook. Reproducing classic factors is a useful exercise: turning paper formulas into computable signals, understanding market-wide stock rankings, and then testing predictive power, stability across periods, and returns after costs. Even without finding a strategy ready for use, the process has learning value. By reproducing these formulas myself, I want to understand what information they offer in the Japanese market and where their limitations lie.
The findings first: have classic quant strategies become outdated?
These results do not support directly adopting Alpha101 as an effective stock-selection strategy for Japanese equities today. We could compute 100 formulas, and 50 theoretical portfolios had positive average returns after costs, but none outperformed TOPIX on the same basis. The number of positive Rank IC results that remained significant after correction for multiple testing was also zero.
Being classic, public, and widely cited is therefore insufficient reason to use a set of formulas. For investors considering a direct application of Alpha101, this test offers insufficient support: the formulas generated signals, but those signals did not translate into a demonstrable stock-selection advantage after costs during this validation period.
These results also do not establish that classic quant strategies as a whole are outdated. The test covers public formulas applied to Japanese equities over a particular period, with a five-day holding period and costs of 25 bps per side. These results alone cannot disentangle market differences, decay in an advantage, and the effects of trading costs. The question here is whether a classic set of formulas can pass a fresh test in a new market and period, rather than being adopted on reputation alone.
Using the January–July 2026 historical validation period and returns over 5 trading days as the main evaluation framework:
| Question | Count |
|---|---|
| Computable formulas out of 101 | 100 |
| Positive average Rank IC | 24 |
| Positive Rank IC and individually significant, unadjusted p<0.05 | 2 |
| Positive Rank IC and still significant after correction across 101 tests | 0 |
| Positive average theoretical upper-quantile portfolio returns after 25 bps per side | 50 |
| Net returns above TOPIX for the corresponding intervals, on the same cost basis | 0 |
The formula we could not compute was #56, because reliable daily point-in-time market capitalization was unavailable.
IC measures the correlation between a factor and subsequent returns; Rank IC compares stock rankings. A positive Rank IC means stocks with higher factor values tend to earn higher returns. The small number of individually significant results no longer remained significant when accounting for the simultaneous testing of 101 formulas. They do not justify declaring that an effective strategy has been found.
Across periods, 14 formulas passed the main criteria—including coverage, positive Rank IC, and significance—in the second half of 2025; none passed in January–July 2026. Average Rank IC changed from positive to negative for 60 formulas. Stability is harder to establish than apparent effectiveness over one period.
50 positive returns did not become 50 excess returns. A rising market can make a portfolio profitable without demonstrating stock-selection skill. After costs of 25 bps per side, no formula's theoretical portfolio had average net returns above the TOPIX price returns corresponding to its valid portfolio periods.
These comparisons after costs are historical theoretical portfolio results, not actual account profits. The averages include only portfolio periods with complete entry and exit data. They do not establish that every formula underperformed the index over the full interval. The validation period is also not a pristine, previously unused out-of-sample period.
What is Alpha101?
Alpha101 is a set of 101 quantitative factor formulas published in the paper 101 Formulaic Alphas. They combine prices, trading volume, returns, VWAP, industry classifications, and other inputs into numerical values for ranking stocks. Some are simple; others layer correlations, rolling ranks, and industry neutralization.
In the direction specified by the original formulas, higher factor values generally suggest a more favorable view of a stock. But a factor is not a complete strategy: the formulas do not also answer questions about position sizing, execution costs, or risk control. Nor does “Alpha101” mean that all 101 formulas can generate excess returns in every market.
Why try them in Japan?
We wanted to test whether these price and volume relationships transfer to Japanese equities, rather than simply accepting the effectiveness of public formulas. Japan offers another stock universe in which to examine which relationships might transfer across markets and which may depend on a particular market environment.
These formulas primarily use market data, making them a useful common starting point for Japanese equity factor research. With roughly two years of Japanese equity tick data available, we could compare the full set rather than selecting a few seemingly profitable examples. The aim was to find relationships that consistently help select stocks, rather than portfolios that happen to make money when the market rises.
Two actual market-wide cross-sectional examples
The following weak-signal examples are chosen specifically for explanation and do not represent all research candidates. Both use actual J-Quants market data from 2026-04-01 to show the formulas, cross-sectional rankings, and returns of five groups. Their performance cannot substitute for the conclusions of the full study. R denotes the cross-sectional percentile rank in that day's base stock universe; final portfolios use stocks meeting the research criteria.
#30: three days of price direction and the share of recent volume
Here, C₁, C₂, and C₃ are the closing prices 1, 2, and 3 trading days earlier. sign assigns +1 to a rise, −1 to a fall, and 0 to an unchanged price. ΣV₅ and ΣV₂₀ are cumulative trading volumes over the 5 and 20 days ending on the observation date.
Using Toyota (7203) as one worked row, the actual closing prices were:
| Date | Closing price (JPY) |
|---|---|
| 2026-03-27 | 3,408.00 |
| 2026-03-30 | 3,224.00 |
| 2026-03-31 | 3,162.00 |
| 2026-04-01 | 3,311.00 |
The sum of the directions of the three consecutive close-to-close changes is -1. Its rank in that day's base stock universe is 0.2968834688. The 5-day volume total is 127,508,000 shares, and the 20-day total is 448,072,700 shares, giving:
This Toyota row simply illustrates how to substitute inputs into the formula. Grouping and IC must use all valid stocks, with the following results.
The valid cross-section for #30 that day contained 1,829 stocks. Stocks were divided into Q1–Q5 from low to high factor values, and subsequent equal-weight returns were calculated for each group, with entry on 2026-04-02 and exit on 2026-04-08.
| Group | Stock count | Average 5-day gross return |
|---|---|---|
| Q1 | 365 | +2.4531% |
| Q2 | 366 | +2.0900% |
| Q3 | 366 | +3.3470% |
| Q4 | 366 | +3.1333% |
| Q5 | 366 | +2.2735% |
The cross-sectional Rank IC that day was +0.002440. Q5's theoretical net return after 25 bps per side was +1.7678%. This is the subsequent-period result for one day's cross-section, not performance over the entire validation period or actual account performance.
Figure: Each point represents one stock; the five group bars show equal-weight returns for the full valid stock universe. This fixed single-day example does not represent long-term performance.
Supporting files: #30 rankings, groups, and returns for all stocks, step-by-step formula calculation fields.
#72: a ratio of ranks for two price–volume relationships
The paper's formula is:
In the paper, adv40 means 40-day average trading volume. This study's implementation uses 40-day average daily traded value in yen, so the reproduction below is of that implementation and cannot be treated as fully equivalent to the paper's original formula. ts_rank ranks observations within a stock's rolling window; decay_linear assigns higher linear weights to more recent observations. Following the study's common convention, fractional windows are rounded down, giving 8, 10, 3, 18, 6, and 2 days in sequence.
For the Toyota row, the numerator follows “8-day correlation between the intraday midpoint price and 40-day average traded value → 10-day linear weighting → cross-sectional rank across all stocks.” The denominator follows “6-day correlation between the 3-day rank of VWAP and the 18-day rank of volume → 2-day linear weighting → cross-sectional rank.” The actual values are:
| Calculation step | Numerator path | Denominator path |
|---|---|---|
| Rolling correlation | -0.6870717756 | -0.5672816854 |
| Value after linear weighting | -0.2584221930 | -0.5687775857 |
| Cross-sectional rank in that day's base stock universe | 0.3624331174 | 0.2097272983 |
All these values match the saved factor values. The complete rolling calculation fields are retained in the supporting files, and stock rankings and grouping also use the full cross-section.
The valid cross-section for #72 that day contained 1,813 stocks. Stocks were divided into Q1–Q5 from low to high factor values, and subsequent equal-weight returns were calculated for each group, with entry on 2026-04-02 and exit on 2026-04-08.
| Group | Stock count | Average 5-day gross return |
|---|---|---|
| Q1 | 362 | +3.1422% |
| Q2 | 363 | +2.3535% |
| Q3 | 362 | +2.3422% |
| Q4 | 363 | +2.4616% |
| Q5 | 363 | +3.1579% |
The cross-sectional Rank IC that day was +0.012383. Q5's theoretical net return after 25 bps per side was +2.6500%. This is the subsequent-period result for one day's cross-section, not performance over the entire validation period or actual account performance.
Figure: Each point represents one stock; the five group bars show equal-weight returns for the full valid stock universe. This fixed single-day example does not represent long-term performance.
Supporting files: #72 rankings, groups, and returns for all stocks, step-by-step formula calculation fields.
Conclusion: classic formulas still need fresh tests
If the question is whether Alpha101 can be directly adopted today to gain a stock-selection advantage in Japanese equities, these results say the evidence is insufficient and does not justify direct use. Under the five-day validation framework for January–July 2026, 50 of the 100 computable formulas had positive average theoretical portfolio returns after costs, but zero had positive net excess returns over TOPIX. Zero also had positive Rank IC that remained significant after correction across 101 tests. Making money and demonstrating effective stock selection require separate judgments.
If the question is whether classic quant strategies as a whole have become outdated, this study is insufficient to make that judgment. Alpha101 is a set of public factors and cannot represent every quantitative method. This study also does not provide a complete body of evidence that these factors once worked and now fail in the same Japanese market under the same cost framework. Performance in the discovery period followed by failure to pass validation shows that stability needs retesting; it does not by itself establish that strategy aging caused the failure.
The clear implication for research and use is: Alpha101 remains useful as a research baseline, but this validation provides no basis for putting it directly into use. The value of classic formulas must be demonstrated again through current evidence of predictive power across periods and excess returns after costs. The two single-day examples only explain how formulas produce market-wide rankings. A positive return or IC on one day does not change that conclusion.
As an entry in my workbook, the value of this reproduction goes beyond whether I found a profitable factor. Classic formulas offer concrete exercises with results that can be checked. They helped me connect formula implementation, cross-sectional comparisons, statistical testing, and cost assessment, and distinguish “computable,” “contains predictive information,” and “can generate excess returns after costs.” Reproducing classic factors remains a useful starting point for learning. The value of the exercise and a strategy's readiness for direct use are separate questions.
Formula source: 101 Formulaic Alphas, original paper.