1. The trading rule almost writes itself
A fear-and-greed index invites a trading rule before it invites a research question. Buy when everybody seems frightened. Wait for confidence to return. Sell when optimism becomes excessive. The appeal is easy to understand: it promises a disciplined way to do what investors often find emotionally difficult.
Our first study of Japanese equity market sentiment found more support for interpreting J-FGI as a description of market risk state than as a directional-return predictor. That leaves the intuitive trading idea open. A weak relationship with a fixed future return does not tell us exactly what happens when the exit depends on sentiment itself.
J-FGI (Japan Fear & Greed Index) is the Japanese equity sentiment index I constructed from four equally weighted dimensions: Trend, Market Strength, trading-value Breadth and Realized Stress. It is my research indicator, rather than an official Japanese sentiment index or a Japanese edition of CNN's Fear & Greed Index. This article uses the definition and scores fixed in the first study.
We therefore kept one primary rule: enter when J-FGI is at or below 20, and exit when it is at or above 80. Does that simple rule survive a historical test? More precisely, does it produce returns that look unusual after allowing for how long we waited?
Figure 1. Four complete primary 20→80 cycles in the frozen Core observation window. Entry and exit markers indicate the next actual trading-day close after each signal; the J-FGI and TOPIX panels have separate scales.
2. How the historical test works
Here, a trading day means a day on which the exchange actually opens, excluding weekends and market holidays. Holding duration counts the trading-day intervals between the purchase and sale dates: the first trading day after purchase is day 1. A holding period of 289 trading days therefore does not mean 289 calendar days.
The rule moves between cash and one long index position. While in cash, the first qualifying Fear observation generates an entry signal. While invested, the first qualifying Greed observation generates an exit signal. More Fear observations during a holding period do not add positions. The exit signal day does not also trigger a new entry.
Observation and execution are separate: a signal observed on day t is assigned the closing price on the next actual trading day. We use that convention to calculate each holding period's price change.
The experiment fixes the entry and exit rules first, then records each cycle's return, holding duration and drawdown.
3. Why TOPIX and a Nikkei proxy
TOPIX supplies the broad Japanese equity market reference. For the second exposure, we use the NEXT FUNDS Nikkei 225 ETF (1321) proxy. Throughout this article, “1321” means that ETF price series; it does not mean the Nikkei 225 cash index.
The local 1321 closing-price series is corrected using cumulative adjustment factors. It does not reinvest dividends. Fund expenses and tracking differences remain reflected in its market prices. TOPIX is also a price index, excluding dividends. Both sets of results therefore describe price changes, rather than an investor's dividend-inclusive total return.
Using two exposures helps reveal whether the broad conclusion depends on the chosen market representation. It does not double the number of independent sentiment episodes: the signals occur on the same dates. Differences between the two return columns are informative, but they are not a second independent experiment.
4. The first result looks encouraging
The primary rule completed four cycles for each asset. The ETF proxy finished positive three times; TOPIX finished positive twice. On a first reading, the average returns might make the familiar instruction to buy fear sound sensible.
| Primary 20→80 result | TOPIX | 1321 ETF proxy |
|---|---|---|
| Complete cycles | 4 | 4 |
| Win rate | 50% | 75% |
| Mean cycle return | +4.01% | +7.39% |
| Median cycle return | −0.30% | +4.74% |
| Median holding period | 212.5 trading days | 212.5 trading days |
These are arithmetic summaries of individual cycles, not annual returns or the compounded performance of a complete portfolio. The difference between TOPIX's positive mean and negative median also tells us to look closely at the individual outcomes.
The holding-period row deserves equal attention. An average gain acquired over a few weeks and one acquired over more than a year answer different investment questions. Without examining that time exposure, “the trades made money” is an incomplete result.
5. The individual cycles matter more than the average
The first TOPIX cycle entered on October 17, 2014 and exited on December 3. Its +21.45% return arrived in just 31 trading days. That is the kind of rebound people remember when they describe buying panic.
The next cycle offers a different experience. An August 25, 2015 entry remained open until November 1, 2016: 289 trading days, a −2.75% final return, and a −25.51% maximum drawdown along the closing-price path. Maximum drawdown measures the fall from a running peak; it need not equal the loss relative to the initial purchase price.
The February 7, 2018 entry took 420 trading days to reach its October 30, 2019 exit. TOPIX finished down 4.80%. A rule that waits for sentiment to reach Greed can require considerable patience without rewarding it.
Finally, the February 26, 2020 entry finished up 2.16% after 136 trading days. During the holding period, both its maximum drawdown and its maximum adverse closing-price move from entry were −23.03%. An eventual small profit required enduring a substantial decline first.
“Eventually profitable” and “useful timing rule” are different claims. These path statistics use daily closes, so they do not describe the worst intraday experience.
Figure 2. Individual primary-cycle price returns. TOPIX returned +21.45%, −2.75%, −4.80% and +2.16%; 1321 returned +21.74%, −1.65%, +5.71% and +3.76%, in entry-year order.
Figure 3. Final return and the path taken to reach it. The 2020 TOPIX cycle ended at +2.16% after a −23.03% maximum drawdown; the corresponding ETF proxy ended at +3.76% after −26.12%.
6. The harder question: compared with what?
Suppose we hold Japanese equities for 289 trading days. Even without a sentiment signal, there is a historical distribution of returns for that length of exposure. Comparing a Fear entry only with holding cash cannot tell us how much of its outcome belongs to the entry rule and how much belongs to simply owning equities for a long time.
For each actual cycle, the study enumerates other historical starting dates for the same asset and exactly the same holding duration, requiring a complete positive-price path and excluding the actual entry itself. It then places the cycle return within that distribution.
A percentile near 50 means the return sits around the middle of those matched windows. A percentile near 100 means it exceeded almost all of them. It is a historical rank, not a probability of future success. The full-history comparison also includes periods after the entry, making it an explicitly retrospective benchmark rather than information an investor necessarily had at the time.
7. Matched-duration results change the story
The 2014 rebound stands out in both assets. The later cycles tell a much less consistent story.
| Entry year | Holding trading days | TOPIX percentile | 1321 percentile |
|---|---|---|---|
| 2014 | 31 | 99.89 | 99.87 |
| 2015 | 289 | 28.63 | 23.03 |
| 2018 | 420 | 20.70 | 24.94 |
| 2020 | 136 | 44.40 | 47.80 |
These are full-history matched-duration percentiles. In three of the four cycles, both assets fell below the midpoint of their respective distributions. Even the positive 2020 outcomes were unexceptional against other windows with the same time exposure.
Figure 4. Full-history matched-duration percentiles for the primary rule. Only the 2014 cycle approaches the top of the historical distribution; the other three sit below its midpoint in both assets.
The comparison pools are not identical across assets because their available price histories differ. Historical starting windows overlap, too: thousands of candidates do not amount to thousands of independent observations. Other pool definitions can move the ranks, so the table should be read as a particular counterfactual, not a universal score.
A memorable successful rebound can make the rule look much stronger than the complete set of cycles does. The matched comparison brings that first success back into proportion.
8. Random starts give the same warning
The study also randomly samples one matched-duration historical start for each actual cycle and averages the sampled returns. Repeating that exercise 1,000 times creates a descriptive reference distribution for the mean of four cycle-length exposures.
TOPIX's actual mean was +4.01%, compared with a random-start mean of +7.95%. Its actual mean occupied the 35.3rd percentile of the sampled distribution. For 1321, the corresponding figures were +7.39%, +11.51% and the 32.1st percentile.
Figure 5. TOPIX primary mean against 1,000 random-start replicates with the same four holding durations. The reference mean is +7.95%; the actual +4.01% mean is at percentile 35.3. This is a descriptive control, not a significance test or portfolio simulation.
The Fear→Greed average did not occupy an unusually strong position relative to randomly sampled historical starts with the same holding durations. This comparison describes return ranks, rather than an alternative trading strategy.
9. Does changing the thresholds rescue the rule?
Alongside 20→80, the study specified three robustness variants in advance: 20→70, 30→80 and 30→70. Earlier exits or less extreme entry conditions naturally change the events and the time spent invested.
| Plain rule | Cycles per asset | TOPIX median return | 1321 median return |
|---|---|---|---|
| 20→80, primary | 4 | −0.30% | +4.74% |
| 20→70 | 6 | +2.83% | +7.48% |
| 30→80 | 6 | +1.79% | +5.94% |
| 30→70 | 9 | +4.10% | +4.30% |
Some summaries look better. The 1321 30→80 variant, for example, had an 83.33% win rate. But all four rules still contain only four to nine complete cycles per asset. These neighboring definitions explore sensitivity; they do not supply enough evidence to select a winner.
Figure 6. Robustness across entry thresholds 20 and 30 and exit thresholds 70 and 80. Each cell reports cycle count and median price return; 20→80 remains the primary rule regardless of the historical ranking.
Robustness analysis stops being robustness analysis when the best historical variant becomes the new rule. That would answer a different question from the one we set out to test.
10. What about waiting for fear to start recovering?
A secondary hypothesis adds one condition: enter during Fear only when the frozen five-day change in J-FGI is positive. The intuition is to wait for sentiment to begin recovering instead of buying at its first low reading.
Under 20→80, that condition raised the mean to +5.46% for TOPIX and +8.97% for 1321. Each result, however, rests on just three complete cycles. Across the pre-specified thresholds, the secondary sets contain only three to six cycles.
The filter changes both entry timing and the set of qualifying events. We are therefore not comparing identical trades with one small adjustment, and the two sets are not independent randomized samples. A better mean could reflect which episodes remain, when they enter, or a small number of favorable outcomes. The hypothesis is interesting enough to keep studying, nowhere near enough to declare an improvement.
11. Fear does not imply an immediate rebound
There is another way to separate the entry idea from the Greed exit: start at the first qualifying observation in each continuous Fear episode, then measure a fixed holding horizon. This produces a different event set from the state-machine cycles. The plain extreme-Fear summaries contain 36 observations per asset at each displayed horizon.
| Trading days after entry | TOPIX mean return | 1321 mean return |
|---|---|---|
| 1 | +0.08% | +0.10% |
| 5 | −0.08% | +0.10% |
| 10 | +0.40% | +0.66% |
| 20 | +1.20% | +1.41% |
| 60 | +3.02% | +4.42% |
There is little here to support a dependable immediate rebound. The longer horizons become more positive, but ordinary equity drift can also contribute to positive returns over longer periods. These means alone cannot assign the gain to sentiment timing.
The events and their future windows can overlap. Nor should identical Fear-entry observations repeated under different Greed specifications be added together to enlarge the sample. Decomposing the rule helps clarify the mechanism; it does not make the evidence independent.
12. Why this is not a trading strategy yet
The Core observation window runs from July 31, 2014 through September 30, 2020. A longer price history supports matched-duration comparisons; it does not turn this into an eighteen-year history of observable sentiment cycles.
The underlying historical inputs are current-vintage. We have not established when the records were published or available at the time, so the study cannot yet reconstruct the information investors could actually have seen. Assigning execution to the next trading day does not supply the missing timestamps. Point-in-Time Data Is Not a Boolean examines this distinction: a past value we can calculate today may not have been available then.
Calendar and historical-universe issues remain unresolved, and comparison windows overlap. Commissions, bid–ask spreads, capacity and intraday execution are not modeled; there is no live or forward validation. A next-day closing price does not establish signal availability before an order deadline or a fill at that price. Neither series includes dividend reinvestment, and 1321 remains an ETF proxy. Four primary cycles per asset are too few to establish a reliable Fear-entry effect.
This backtest describes historical price paths under fixed rules. The next steps are to verify historical information availability, include execution costs and test the rule on new observations—the checking process discussed in Why ArcaViso and the KabuForge research framework.
13. Fear may describe a state without creating an edge
The first article asked what J-FGI measures. This follow-up asks why that interpretation matters when an appealing label becomes an investment rule.
A stressed market state can be economically meaningful without offering an exceptional entry price for the holding period that follows. In this sample, one outstanding rebound coexists with long waits, substantial adverse moves and three later cycles that rank below the full-history matched-duration midpoint in both assets.
Assessing a Fear→Greed timing edge requires looking at return, holding duration and losses along the way together. In this matched-duration comparison, only the 2014 rebound exceeded its historical comparison midpoint; the other three primary cycles did not.
Fear and greed may be useful names for market states. That does not automatically make them trading signals.
Construction and sources
- Original research: I constructed J-FGI and the Fear→Greed experiment. Definitions, thresholds and samples remain those of the fixed study. Broad Risk-On, from my personal trading system, is a market-state comparison in the first article rather than an external signal source here.
- Market data: Prices and calendars come from locally saved J-Quants historical records. See JPX J-Quants and the official API client and documentation. Historical availability of the study's data versions remains unverified.
- Asset identity: NEXT FUNDS' official 1321 information identifies the Nikkei 225-linked ETF. Returns here are calculated from the local price series, rather than taken from the fund's published performance.
- Conceptual reference: CNN Fear & Greed Index provides an example of composite sentiment scoring. J-FGI's four-dimensional Japanese market construction and this trading experiment are my own research.