Methodology

This page shows, without hiding anything, how we validate our rules and what their limits are.

'Market now' describes the current state, not a forecast

The old verdict tried to say whether the market would be up or down a month later. Tested fairly on past data the model had not seen, its balanced accuracy was 48-49% in all four markets (Bitcoin 47.9%) - about a coin flip.

We also built a model to forecast next month's swings. It did no better than simply saying 'the current state continues' (difference -4.4 to +2.0 pts, not statistically distinguishable).

So from 2026-09-22 the screen is not a forecast. It describes how much the market is swinging now compared with usual. No hit rate is attached.

  • Quiet - recent swings below 0.8x usual (clears at 0.9x or above).
  • Normal - moving about as usual.
  • Volatile - recent swings at 1.3x usual or more (clears below 1.15x).
  • Falling sharply - more than 12% below the 60-day high with swings at 1.4x usual or more, or a one-day drop of more than 4% (and more than 4x a usual day) in the last 5 sessions. KOSDAQ uses 18%; Bitcoin uses 20% and 8% a day.

Swings = standard deviation of daily log returns over the last 20 trading days. Usual = the median of that value over the past year (365 days for Bitcoin). To avoid flip-flopping, the bar to leave a state is looser than the bar to enter it.

Applied to history (US since 1991, KOSPI 1998, KOSDAQ 2001, Bitcoin 2015), the state changes about once a month. The month after a quiet day swung about 0.8x usual; after a volatile day about 1.2x. It is a factual description that recent swings tend to last a few weeks.

It does not say whether prices will rise or fall. Not investment advice.

How the previous method (direction forecast) tested

MarketBalanced accuracy, up & down days (%)Always saying 'up' (%)Swing forecast − simple persistence (pp, 95% range)
US49.364.9+0.4 (-2.6 ~ +3.1)
KOSPI49.157.8+1.1 (-2.2 ~ +4.2)
KOSDAQ48.854.8-4.4 (-8.3 ~ -0.5)
Bitcoin47.952.3+2.0 (-4.1 ~ +7.2)

out-of-sample, yearly re-fit, 21-trading-day return sign; US 1995-2026, KOSPI 2002-2026, KOSDAQ 2005-2026, BTC 2018-2026

The current market state applied to history

MarketPeriodQuiet / Normal / Volatile / Falling sharply (% of days)Swings a month later (x usual, same order)Changes per month
US1991-01-28 ~ 2026-09-1835.4 / 37.4 / 20.8 / 6.40.80 / 0.97 / 1.24 / 1.620.91
KOSPI1998-01-30 ~ 2026-09-1831.2 / 42.4 / 17.2 / 9.20.83 / 0.99 / 1.17 / 1.451.03
KOSDAQ2001-11-22 ~ 2026-09-1834.6 / 34.7 / 21.2 / 9.50.83 / 0.99 / 1.21 / 1.291.23
Bitcoin2015-10-06 ~ 2026-09-1935.7 / 35.5 / 17.8 / 11.00.83 / 0.94 / 1.20 / 1.271.07

Record archive

Previous-method (direction forecast) records are kept and still scored by the 21-trading-day rule. Calendar: US · Korea · Crypto · commodity

How we validate

  • Walk-forward: scores use only data available up to the rebalance date, and results are checked only against what actually happened afterward.
  • Out-of-sample testing: performance is measured on periods not used to set the rules.
  • No lookahead: data not yet available at that point in time (e.g. the next day's close) is never used in the calculation.
  • Transaction cost assumptions: minimum validation window 3 years, commission 0.005/share, minimum commission 1.0, slippage 5bp, minimum paper-trading period 90 days.

Survivorship bias

Backtesting only with stocks that are alive today inflates results, since delisted stocks are excluded. So the goal is to store the universe (candidate list) as a weekly snapshot of what actually existed at that time. The buy-fitness table doesn't have that snapshot yet (partial coverage), so it uses today's universe even for past periods — stocks that were delisted or newly added in between aren't reflected.

Rules & parameters (fixed in advance)

The numbers below were fixed when the strategy was designed, not fitted after looking at the data. By default, buy candidates exclude companies with a caution signal in revenue, profit, or debt (e.g. ongoing losses, high debt) — this can be turned off per strategy.

Parametervalue
top_k20
horizon_days21
max_position_pct10.0%
weight · trend_1201.0
weight · momentum_12_11.0
weight · zscore_20-0.5
weight · vol_regime-0.5
weight · dist_high_2520.5
exit.balancedstop 8% · trail 10% · take-profit 25% · max 120d
exit.trendstop 10% · trail 15% · take-profit 40% · max 180d
exit.dipstop 6% · trail 8% · take-profit 15% · max 60d
exit.stablestop 5% · trail 7% · take-profit 12% · max 90d

Parameter sensitivity

We check that small changes to the stop-loss level or the trend lookback window don't swing the results wildly (an overfitting check).

US

ParameterAvg returnHit rateSample
trend_120 = 100d+1.87%56%107
trend_120 = 120d+2.09%57%107
trend_120 = 150d+2.27%57%107
stop = 6%+0.73%46%2140
stop = 8%+1.04%51%2140
stop = 10%+1.18%53%2140

Korea

ParameterAvg returnHit rateSample
trend_120 = 100d+2.06%48%104
trend_120 = 120d+2.28%49%104
trend_120 = 150d+2.25%48%104
stop = 6%+1.09%33%2080
stop = 8%+1.42%38%2080
stop = 10%+1.81%43%2080

Crypto

ParameterAvg returnHit rateSample
trend_120 = 100d+2.70%44%101
trend_120 = 120d+2.81%44%101
trend_120 = 150d+2.69%44%101
stop = 6%+2.57%29%1967
stop = 8%+2.50%31%1967
stop = 10%+2.65%34%1967

Buy fitness

In past walk-forward validation, the share of stocks that rose 21 trading days (about a month) later when a (market verdict stage) × (that day's rank bucket) × (caution flag present/absent) condition held. No arbitrary weights. Verdict-stage labels are today's rules applied retroactively to the whole history (an approximation — those rules didn't necessarily exist back then).

US

Verdict stageRank bucketCaution flagShare upAvg returnSample (stock-days)Sample (trading days)
Risk1-5No57.5%+3.00%749167
Risk1-5Yes53.5%+5.73%8657
Risk6-10No54.7%+2.32%697167
Risk6-10Yes53.6%+0.69%13891
Risk11-20No57.8%+2.43%1387167
Risk11-20Yes58.0%+1.85%283115
Risk21-50No56.8%+1.71%3988167
Risk21-50Yes56.9%+2.78%1022157
Caution1-5No53.5%+0.95%53001210
Caution1-5Yes53.7%+0.76%775483
Caution6-10No53.5%+0.68%49561211
Caution6-10Yes54.9%+0.85%1119656
Caution11-20No53.3%+0.65%96311215
Caution11-20Yes56.6%+0.94%2519915
Caution21-50No53.7%+0.69%283581215
Caution21-50Yes54.7%+0.64%80921153
Mixed1-5No63.7%+3.38%3546764
Mixed1-5Yes57.7%+2.31%284184
Mixed6-10No60.4%+2.75%3425763
Mixed6-10Yes58.8%+2.48%405261
Mixed11-20No60.1%+2.54%6685766
Mixed11-20Yes58.3%+2.39%975402
Mixed21-50No60.1%+2.41%19934766
Mixed21-50Yes57.7%+1.41%3046582
Good1-5No70.1%+4.33%40283
Good1-5Yes68.9% (Small sample · merged with a broader condition)+4.04%41583
Good6-10No62.9%+2.92%38083
Good6-10Yes60.0%+2.90%3528
Good11-20No63.7%+3.68%75883
Good11-20Yes66.7%+4.80%7240
Good21-50No67.9%+4.60%227883
Good21-50Yes55.7%+1.38%21264

Korea

Verdict stageRank bucketCaution flagShare upAvg returnSample (stock-days)Sample (trading days)
Risk1-5No44.7%+2.86%9420
Risk1-5Yes45.0% (Small sample · merged with a broader condition)+2.80%10020
Risk6-10No51.1%+3.56%9220
Risk6-10Yes52.0% (Small sample · merged with a broader condition)+3.77%10020
Risk11-20No46.0%+4.83%17420
Risk11-20Yes48.0% (Small sample · merged with a broader condition)+5.24%20020
Risk21-50No49.1%+2.37%50720
Risk21-50Yes50.5%+6.38%9319
Caution1-5No47.8%+0.86%1501328
Caution1-5Yes55.4%+7.47%139111
Caution6-10No45.6%+0.60%1415328
Caution6-10Yes48.0%+4.72%225162
Caution11-20No47.0%+1.19%2812328
Caution11-20Yes47.4%+2.28%468250
Caution21-50No45.3%-0.12%8543328
Caution21-50Yes44.1%+1.61%1285309
Mixed1-5No47.7%+1.74%48421073
Mixed1-5Yes54.0%+7.21%517361
Mixed6-10No46.7%+2.13%46401070
Mixed6-10Yes48.2%+3.28%710496
Mixed11-20No48.9%+1.93%90051070
Mixed11-20Yes46.9%+2.22%1685768
Mixed21-50No47.5%+0.77%267741068
Mixed21-50Yes48.8%+2.17%52201011
Good1-5No51.3%+3.54%3381743
Good1-5Yes58.7%+4.27%334254
Good6-10No54.0%+3.72%3214739
Good6-10Yes57.5%+5.54%501326
Good11-20No54.6%+3.87%6328743
Good11-20Yes52.0%+4.02%1102509
Good21-50No55.5%+3.64%18634743
Good21-50Yes53.1%+4.70%3656668

Crypto

Verdict stageRank bucketCaution flagShare upAvg returnSample (stock-days)Sample (trading days)
Risk1-5No36.2%-2.62%18841
Risk1-5Yes37.1% (Small sample · merged with a broader condition)-1.94%20541
Risk6-10No29.0%-4.72%18341
Risk6-10Yes29.3% (Small sample · merged with a broader condition)-4.44%20541
Risk11-20No31.7%-6.45%35741
Risk11-20Yes19.6%-7.21%5114
Risk21-50No13.8%-11.90%14536
Risk21-50Yes13.1% (Small sample · merged with a broader condition)-12.62%16039
Caution1-5No43.3%-0.01%47491256
Caution1-5Yes50.5%+11.45%640366
Caution6-10No43.7%+0.31%3996998
Caution6-10Yes46.2%+3.79%688338
Caution11-20No37.3%-1.63%6424797
Caution11-20Yes38.6%+2.86%1173335
Caution21-50No27.9%-7.50%1429448
Caution21-50Yes19.2%-10.08%255116
Mixed1-5No52.0%+6.06%79361884
Mixed1-5Yes53.7%+9.11%1380761
Mixed6-10No50.1%+4.95%72571768
Mixed6-10Yes52.5%+9.22%1270656
Mixed11-20No48.3%+2.77%104161372
Mixed11-20Yes51.7%+10.01%1629568
Mixed21-50No39.5%-1.89%2632802
Mixed21-50Yes46.2%+4.18%379200
Good1-5No––00
Good1-5Yes––00
Good6-10No––00
Good6-10Yes––00
Good11-20No––00
Good11-20Yes––00
Good21-50No––00
Good21-50Yes––00

Sample (stock-days) is the count of stock-date pairs meeting the condition; sample (trading days) is the number of distinct trading days among them. If either falls short of the minimum (30 stock-days, 10 trading days), we show '–' instead of a value.

Testing the company-financials filter

We re-ran the rule 'drop companies with risky financials from the candidate list' using only the figures that were already filed on each day (point-in-time — no later restatements, nothing not yet published). For every trading day we take the 20 candidates and compare their return 21 trading days later against the index. Because holding periods overlap, days are not independent, so the 95% interval comes from a 21-day block bootstrap (1,000 resamples). The bar for 'adopted' was set before looking at the results: the lower bound must be above 0 and the sign must agree in at least two half-year periods.

US

Sample: 2025-08-11 to 2026-08-06, 249 trading days, 4960 stock-days. Financial verdict available for 100% of candidates.

Financial summary21-day meanShare that rosevs indexCases (stock x day)
Solid+5.16%55%+4.00%732
Average+2.12%58%+0.44%2938
Caution+2.02%51%+0.15%1290
No data–––0
How candidates are pickedvs indexDifference vs default95% intervalVerdict
Default (by score)+0.89%–––
Exclude caution companies+1.27%+0.38%[-0.17%, +1.25%]On hold (not enough evidence)
Solid companies first+1.90%+1.01%[-0.24%, +2.51%]On hold (not enough evidence)

Korea

Sample: 2025-07-28 to 2026-08-06, 250 trading days, 5000 stock-days. Financial verdict available for 100% of candidates.

Financial summary21-day meanShare that rosevs indexCases (stock x day)
Solid+7.61%63%-0.57%826
Average+6.04%56%-1.22%3314
Caution+2.73%55%-4.16%854
No data-11.67%17%-11.97%6
How candidates are pickedvs indexDifference vs default95% intervalVerdict
Default (by score)-1.63%–––
Exclude caution companies-1.37%+0.26%[-0.33%, +0.96%]On hold (not enough evidence)
Solid companies first-1.71%-0.08%[-1.43%, +1.03%]On hold (not enough evidence)

Limits: industry labels and the stock list are today's, applied backwards, so companies that disappeared in between are missing. The test window is about a year, giving only a dozen or so non-overlapping one-month periods - too few for a strong statistical conclusion. Trading costs are not included.

Scoring the sharp-drop risk score

The sharp-drop alert is not a forecast - it is a daily thermometer for stress signs that often show up before big drops. Ten to twelve signals are turned into percentiles (0-1) using only data available up to that day, then combined by a model that is re-trained every January on prior data only, giving a 0-100 score. The stage bands (calm, watch, alert, danger) are chosen on the training window as well, and the table below re-scores only the period that was never used for training. We only have four years of data, so the scored window is short, and outside the US and Japan the danger stage did not separate real drops - treat it as a thermometer, not a signal.

US

Scored window 2019-01-02 - 2026-11-18, 1961 trading days - combined by logit - base rate (max drawdown <= -7% within 21 trading days) 8.9%

Stage (score)Cases (days)-7% rate-10% rateAvg max drawdownvs base
Calm 0-5911085.6%1.8%-2.25%0.63x
Watch 60-794127.5%2.4%-2.22%0.85x
Alert 80-9234915.2%7.7%-2.91%1.71x
Danger 93-1009230.4%16.3%-6.91%3.43x

Of the 12 drops of 10% or more in this window, 9 had the alert at stage 2 or higher within the 5 trading days before.

Korea

Scored window 2003-01-02 - 2026-11-19, 5867 trading days - combined by rank - base rate (max drawdown <= -7% within 21 trading days) 14.4%

Stage (score)Cases (days)-7% rate-10% rateAvg max drawdownvs base
Calm 0-59329611.1%5.1%-3.03%0.77x
Watch 60-79120416.0%8.6%-3.71%1.11x
Alert 80-9297920.9%11.8%-4.02%1.45x
Danger 93-10038821.1%13.4%-4.65%1.47x

Of the 49 drops of 10% or more in this window, 24 had the alert at stage 2 or higher within the 5 trading days before.

Crypto

Scored window 2019-01-01 - 2026-11-19, 2858 trading days - combined by rank - base rate (max drawdown <= -7% within 21 trading days) 38.6%

Stage (score)Cases (days)-7% rate-10% rateAvg max drawdownvs base
Calm 0-59186936.3%25.5%-6.69%0.94x
Watch 60-7949945.1%32.1%-8.02%1.17x
Alert 80-9242641.1%26.1%-6.88%1.06x
Danger 93-1006437.5%25.0%-5.93%0.97x

Of the 93 drops of 10% or more in this window, 35 had the alert at stage 2 or higher within the 5 trading days before.

How to read the statistics

  • If the sample (n) is under 30, chance is too likely a factor — we don't draw conclusions.
  • Don't look at average return alone; check hit rate and sample size together.
  • Past performance does not guarantee future returns.

Limits & regimes where this doesn't work well

  • Being rule-based, it reacts slowly to sudden news or policy shocks.
  • For thinly traded stocks/markets, the slippage assumption may not match reality.
  • Markets with a short history (e.g. crypto) have lower validation confidence.
  • Many days in the forecast record were calculated retroactively, before launch (2026-09-06). Those weren't forecasts we actually published that day, so we exclude them from the scorecard's headline numbers and show them separately, labeled 'backfilled'.
  • The KOSDAQ verdict inside the Korea market (added 2026-09-19) runs the same formula as KOSPI on the KOSDAQ index (^KQ11, Yahoo Finance). The sharp-drop alert is still KOSPI-index-only, and the KOSDAQ forecast record starts 2026-09-19 with no retroactive backfill.

Data sources & license

Price data comes from public market data providers and is provided only as statistical information, not investment advice.

Who makes this

MarketBaro is a side project built by a developer who also runs an AI startup in Singapore. It's not a company product, so we try to keep the rules and the data as transparent as possible.

Every number on this page is a statistic computed by applying fixed rules to past data.

1 2 3