> ## Content Index > Fetch the complete content index at: https://incomequalityscore.com/llms.txt > Use this file to discover other available public pages before exploring further. # Buying and Selling on Candlesticks? Do Not! - URL: https://incomequalityscore.com/buying-and-selling-on-candlesticks-do-not/ - Published: 2026-07-16T08:32:32.000Z - Updated: 2026-07-16T08:32:32.000Z - Description: We tested whether EMA-based buy and sell signals could improve closed-end fund investment outcomes. Across hundreds of CEFs and millions of fund-days, the answer was clear: the data layer worked, but the timing model did not. - Author: Joost Rijlaarsdam - Tags: Research, CEFs, Experiments, EMA, Technical Analysis, Income Quality Score # Buying and Selling on Candlesticks? Do Not! ## Our first CEF timing experiment found that EMA signals do not improve investment outcomes Investors naturally want better timing. Buying a good closed-end fund is one thing. Buying it just before a recovery would be even better. Selling before a decline sounds better still. That is the promise behind technical indicators such as moving averages, trend confirmations and candlestick-based trading signals. They appear objective. They turn a messy price chart into a simple instruction: **Buy. Hold. Sell.** So we decided to test whether that promise holds up for closed-end funds. We built a full historical timing model around exponential moving averages, or EMAs. The model classified CEFs into five ratings: - Strong Buy - Buy - Hold - Sell - Strong Sell We then tested those ratings across hundreds of funds and millions of historical observations. The result was unambiguous: > **The data infrastructure worked. The EMA timing model did not.** Buy signals failed to identify unusually attractive entry points. Strong Buy signals were no better than ordinary Buy signals. Sell signals frequently appeared near market bottoms, shortly before prices recovered. And when we simulated actually following the model, it substantially underperformed simply owning the funds. Our conclusion: > **EMA-based trend following is not a useful timing framework for closed-end funds.** That does not mean price never matters. It means that this specific way of interpreting price—buying confirmed strength and selling confirmed weakness—appears fundamentally mismatched with the behaviour of CEFs. --- ## Why test technical timing for CEFs? Closed-end funds create an interesting timing problem. Unlike ordinary shares, a CEF has two prices: 1. The value of its underlying portfolio, expressed through net asset value. 2. The market price at which investors trade the fund. The market price can move above or below NAV, creating a premium or discount. That introduces opportunities which do not exist in exactly the same form for ordinary stocks. A fund can become temporarily unpopular even when the underlying assets remain sound. Its discount may widen because of sentiment, tax-loss selling, forced liquidation, distribution concerns or general risk aversion. A timing model might therefore help identify when weakness is ending and a recovery is beginning. Our original hypothesis was reasonable: - A bullish EMA crossover might confirm that selling pressure had ended. - Additional confirmation rules might reduce false signals. - Strong Buy signals might identify the highest-quality entry points. - Sell signals might help investors reduce drawdowns. - A five-level rating could provide useful guidance without requiring continuous trading. The point of the experiment was not to prove that technical analysis works. It was to determine whether it works. --- ## The dataset The first major task was building a reliable CEF universe. We examined 22,694 US-listed stock and fund candidates and identified: - **570 confirmed closed-end funds** - **336 active funds** - **234 inactive or delisted funds** - **566 funds with usable price histories** The inactive funds were deliberately retained. Excluding dead or delisted funds would create survivorship bias. It would make the historical universe look healthier than it really was and could exaggerate the success of a timing strategy. Those inactive funds contributed roughly 40% of the available backtest history. In total, the experiment covered approximately: > **2.99 million valid fund-days** This included daily prices, adjusted prices, opening prices, dividends and fund metadata. No second market-data provider was used to repair or supplement the results. For the active universe, 331 of 336 confirmed funds received a valid current rating—a coverage rate of **98.5%**. --- ## Why adjusted prices mattered A CEF’s share price normally drops when it goes ex-dividend. If we had tested moving averages on raw prices, a perfectly normal dividend adjustment could have been mistaken for a bearish market event. That would be a serious problem for high-distribution investments. We therefore validated the adjusted-price history before running the experiment. The adjusted series proved to be an extremely accurate total-return representation. Using raw prices, bearish events clustered more frequently around ex-dividend dates. Using adjusted prices, that false clustering disappeared. This part of the research worked exactly as intended: > Dividend payments did not create artificial sell signals in the final backtest. --- ## The timing model The primary model used a 10-day and 20-day exponential moving average. The basic momentum logic was familiar: - A shorter moving average rising above a longer moving average suggests improving momentum. - A shorter moving average falling below a longer moving average suggests weakening momentum. - Additional confirmation rules were intended to distinguish ordinary signals from stronger ones. - Transitional states attempted to identify bottoms and tops before a full confirmation occurred. The model was tested with several other EMA pairs: - 5 and 13 days - 10 and 20 days - 12 and 26 days - 20 and 50 days This allowed us to determine whether poor results were caused by one unfortunate parameter choice. They were not. Every EMA combination led to broadly the same conclusion. --- ## How the backtest was designed A timing experiment is easy to manipulate unintentionally. We therefore imposed strict point-in-time rules. The model could only use information available on or before the signal date. Signals were executed at the next trading day’s adjusted opening price. The research code was machine-tested to confirm that adding future data did not alter an earlier historical rating. The testing period was divided chronologically: - **Development:** through 2018 - **Validation:** 2019 through 2022 - **Out of sample:** 2023 through July 2026 The primary evaluation horizon—20 trading days—was selected before the backtest results were calculated. That matters because selecting the best-performing horizon afterwards would create another form of overfitting. We also included trading friction: - 30 basis points per round trip in the standard event analysis - Higher-cost stress testing - 15 basis points per transaction side in the portfolio simulation Results were compared against: - Unconditional CEF returns - Randomly selected events - A plain EMA crossover - An EMA-slope signal - Price above the 20-day EMA - Buy-and-hold A complex model should not merely produce positive returns. It should outperform simpler alternatives. This one did not. --- ## The Buy signals did not identify better buying opportunities At first glance, the Buy results might appear respectable. In the out-of-sample period, 60.0% of gross Buy signals were followed by a positive 20-day return. But ordinary CEF days were followed by a positive return **60.1%** of the time. The Buy signal therefore performed slightly worse than the unconditional market base rate. The average gross 20-day forward return tells the same story: | Entry condition | Validation | Out of sample | | --------------------- | ---------- | ------------- | | Full model Buy | +0.36% | +0.64% | | Unconditional CEF day | +0.54% | +0.75% | | Random event | +0.52% | +0.81% | The positive returns after Buy signals were not evidence of timing skill. They were ordinary market drift. CEF prices tend to rise over time because investors receive income and because financial markets have a positive long-term return expectation. A signal should not receive credit merely because prices often rise after it. To add value, the signal must select days that perform better than normal days. The EMA Buy rating did not. --- ## Strong Buy was not stronger The five-level rating system assumed that additional confirmation would improve signal quality. It did not. In validation: - Gross Buy win rate: **62.4%** - Gross Strong Buy win rate: **61.9%** Out of sample: - Gross Buy win rate: **60.0%** - Gross Strong Buy win rate: **60.2%** The differences were economically meaningless. Strong Buy signals did not consistently produce higher returns, higher win rates or better risk-adjusted outcomes than ordinary Buy signals. In some periods they were worse. That is important because a rating system should discriminate. A Strong Buy should represent a materially better opportunity than a Buy. If the tiers do not correspond with increasingly attractive outcomes, the labels create confidence without information. --- ## The Sell signals were not merely weak. They were inverted. The most damaging result came from the bearish side of the model. A Sell signal was considered correct when the fund subsequently declined. In the out-of-sample period, the Sell rating was correct only **39.8%** of the time. Strong Sell was correct only **38.1%** of the time. Put differently: > After an out-of-sample Strong Sell signal, the CEF price rose during the following 20 trading days in approximately 61.5% of cases. The longer the horizon, the more obvious the inversion became. At 60 trading days, the average return following an out-of-sample Strong Sell signal was: > **+4.5%** The Strong Sell win rate at that horizon fell to just **27.3%**. These signals were not identifying market tops. They were frequently identifying market bottoms. A few individual examples were dramatic: - CHN received a Strong Sell signal on September 6, 2024, then rose approximately **49%** over the following 20 trading days. - TDF received a Strong Sell signal on September 9, 2024, then rose approximately **36%** over the following 20 trading days. Those are extreme cases, but they illustrate the broader statistical pattern. The model waited for weakness to become sufficiently established, then told the investor to sell after much of the decline had already occurred. When the price reverted, the investor was left in cash. --- ## Confirmation made the signals worse We also compared the complete rating framework with much simpler indicators. At the 20-day horizon: | Model | Validation average forward return | Out-of-sample average forward return | | --------------------- | --------------------------------- | ------------------------------------ | | Full model Buy | +0.36% | +0.64% | | Plain EMA crossover | **+0.94%** | **+1.02%** | | EMA slope only | +0.95% | +0.90% | | Price above EMA20 | +0.85% | +0.80% | | Random events | +0.52% | +0.81% | | Unconditional returns | +0.54% | +0.75% | The confirmation rules did not filter out noise. They filtered out part of the return. Even the plain EMA crossover produced only a small gross advantage over unconditional returns. Its apparent edge was roughly the size of one round trip in transaction costs. After realistic friction, it was economically irrelevant. The sophisticated version was worse than the simple version. --- ## CEFs showed mean reversion, not useful short-term momentum The core assumption behind an EMA model is that trends persist. Once positive momentum is confirmed, further strength should be more likely. Once negative momentum is confirmed, further weakness should follow. For CEFs, the data showed the opposite tendency at this frequency. During validation, days when the 10-day EMA stood above the 20-day EMA produced an average 20-day forward return of approximately: - **+0.38%** when EMA10 was above EMA20 - **+0.82%** when EMA10 was below EMA20 Out of sample: - **+0.58%** when EMA10 was above EMA20 - **+1.07%** when EMA10 was below EMA20 Buying after confirmed strength produced lower subsequent returns than buying during weakness. That is classic mean-reverting behaviour. It also fits the economic structure of closed-end funds. CEF market prices can temporarily move away from the value of their underlying portfolios. Investor fear can widen discounts. Forced selling can push prices below reasonable value. When sentiment normalises, discounts contract and prices recover. An EMA trend model often reacts only after the movement is already underway. For an asset class with meaningful mean reversion, that delay is costly. --- ## The “bottoming” signal worked—but for the wrong reason One part of the model appeared more promising. The state called `bullish_setup_forming` attempted to identify a potential bottom before the bullish trend had been fully confirmed. Its average 20-day return was: - **+1.26%** during validation - **+1.24%** out of sample That was better than the confirmed bullish states. But this did not validate the trend-following model. It contradicted it. The setup state performed better because it bought unresolved weakness before the moving averages had fully turned upward. In other words: > The best bullish state was the least trend-confirmed state. The potentially useful information was not “the trend has turned.” It was “the fund has become weak enough that mean reversion may be favourable.” That is a different investment thesis and needs a different research framework. --- ## The actual portfolio simulation Event studies are useful, but investors do not experience isolated signals. They experience an entire sequence of buying, selling and waiting. We therefore simulated equal-weight portfolios using the ratings. The results were severe. | Strategy | Validation annualised | Out-of-sample annualised | OOS maximum drawdown | | --------------------------------------- | --------------------- | ------------------------ | -------------------- | | Buy-and-hold eligible CEF universe | **+5.6%** | **+12.6%** | −11.9% | | Hold only while Buy-rated, before costs | −5.1% | +0.9% | −22.6% | | Hold only while Buy-rated, after costs | −13.0% | −8.0% | −32.0% | | Exit on Sell, before costs | −2.7% | +6.9% | −16.9% | | Exit on Sell, after costs | −7.6% | +1.9% | −21.4% | Buy-and-hold did not merely win because of transaction costs. It already won before costs. Costs then made the timing strategies much worse. The hold-while-Buy-rated strategy generated approximately 58 to 61 one-way portfolio turns per year. That added roughly 8% to 9% of annual trading friction in the simulation. Instead of reducing risk, the model also increased drawdowns. Why? Because it repeatedly followed this pattern: 1. Hold the fund during the early part of a decline. 2. Sell after the decline becomes technically confirmed. 3. Remain out during the recovery. 4. Buy again after the recovery becomes technically confirmed. That is close to the opposite of what a long-term CEF investor wants to do. --- ## The ratings changed far too often A useful rating should be stable enough to support decisions. The primary model generated approximately: > **894,000 rating transitions across 2.99 million valid fund-days** That equates to one rating change every **3.3 trading days per fund** on average. Even after excluding certain cooldown events, the average remained roughly one change every 4.3 trading days. In one extreme historical example, the model generated 68 actionable rating changes for a single CEF in just one year. It switched from Buy to Strong Sell in five days and back to Buy one week later. That is not decision support. It is noise translated into labels. The lack of discrimination was also visible in the current snapshot. At one point, 160 of 331 rated funds were simultaneously classified as Strong Buy. When nearly half the universe receives the highest rating, the rating is not meaningfully selective. --- ## Could different EMA settings rescue the model? We tested faster and slower EMA combinations. None changed the conclusion. The 5/13, 10/20, 12/26 and 20/50 combinations all produced broadly unhelpful buy signals and inverted sell signals. Faster moving averages created even more churn. The 5/13 configuration generated approximately 1.17 million rating transitions, compared with roughly 271,000 for the slower 20/50 combination. The slower versions traded less, but they did not discover a reliable edge. This tells us the problem was structural rather than parametric. It was not a matter of finding the magical moving-average pair. The underlying momentum premise was wrong for the task. --- ## What this experiment does—and does not—prove The experiment does not prove that every form of technical analysis is useless for every security. It tested a specific question: > Can EMA-based trend and confirmation signals provide economically useful buy and sell timing for US closed-end funds? The answer was no. More specifically: - Buy ratings did not outperform ordinary CEF exposure. - Strong Buy did not outperform Buy. - Sell ratings were systematically late. - Strong Sell was even more strongly inverted. - Confirmation logic degraded simpler crossover signals. - Alternative EMA periods did not rescue the framework. - Trading costs made an already weak strategy substantially worse. - Drawdowns increased rather than decreased. - The rating system changed too frequently to be useful as an informational product. The proper conclusion is therefore not that markets can never be timed. It is that: > **This signal family should not be productionised for CEF investing.** --- ## What we learned A failed experiment is valuable when it eliminates a bad idea before investors rely on it. The research produced several useful lessons. ### 1\. A positive win rate means little without a base rate A signal that wins 60% of the time sounds impressive. But not when the market also rises 60% of the time without the signal. Investment research must compare every result with the return that was available without taking the action. ### 2\. More confirmation does not necessarily create more certainty Additional rules can make a model look more sophisticated while making its signals later and less useful. In this experiment, confirmation generally meant waiting until more of the price movement had already happened. ### 3\. Selling after weakness is particularly dangerous in CEFs CEF discounts can widen sharply during periods of stress and then mean-revert. A trend model can mistake temporary investor pessimism for a permanent deterioration in value. ### 4\. Price signals cannot be evaluated separately from trading behaviour A model that changes its opinion every few days may look interesting in a chart but become destructive once transaction costs, taxes, spreads and investor behaviour are included. ### 5\. CEF timing probably needs to start with valuation, not momentum The more promising research direction is unlikely to be: > “Has the market price started rising?” It is more likely to be: > “Has the market price become unusually cheap relative to NAV, history, distribution quality and the fund’s underlying condition?” That moves the research away from generic trend following and toward the characteristics that make closed-end funds unique. --- ## The decision We will not use the EMA model as: - a Buy rating; - a Sell rating; - a five-level timing system; - an exit mechanism; - a portfolio risk-control system; - or an informational overlay for CEF investors. The experiment has been stopped. The data infrastructure, adjusted-price validation, historical universe and point-in-time testing framework remain valuable. They can be reused in future experiments. But the model itself will not move into production. That is exactly what disciplined research is supposed to accomplish. Not every experiment needs to uncover a winning strategy. Sometimes the most valuable result is discovering what not to trust. Hopefully, this research also saves you some time. Staring at candlestick charts and watching moving-average crossover lines will not make you a better CEF investor. We tested that idea across hundreds of funds, millions of fund-days and multiple market periods. The evidence simply was not there. But the search does not stop here. CEFs have unusual characteristics—discounts, premiums, distributions and recurring investor overreactions—that may still offer a genuine advantage to disciplined investors. We will keep testing those possibilities until we find signals that survive the data rather than merely looking convincing on a chart. The search for an unfair advantage continues. **We will keep you posted.** --- *Income Quality Score is research based on historical data. It is not personal investment advice and does not tell you what to buy or sell.*