What 2 Million Bitcoin Candles Show About RSI ThresholdsBitcoin / U.S. dollarBITSTAMP:BTCUSDTanHefRSI 30, a bounce? RSI 70, a pullback? Compared with “All candles”, ten candles later: RSI ≤30 events: Ended higher more often on all four timeframes, with the smallest difference on 1h. RSI ≥70 events: Ended higher less often on 5m, 15m and 1h, but more often on 4h. The 4h ≥70 result runs against a simple overbought-reversal expectation. That answers how often price ended higher. Did average returns follow the same pattern? █ How often did price end higher? All results use BITSTAMP:BTCUSD native candles, confirmed-close RSI(14) calculated from closing prices and ten subsequent observed candles. Unless specified, they pool all years in the shared-date cohort defined below. N counts starting closes. For “All candles”, a window starts at every eligible close, including event closes, regardless of RSI. It does not mean RSI 50, crossing 50 or only readings between 30 and 70. It is a historical reference, not an independent control group. Ten candles normally span 50 minutes on 5m, 2½ hours on 15m, 10 hours on 1h and 40 hours on 4h. Gaps can extend elapsed time. How to read these charts Main chart above: Difference from All candles, in percentage points (pp): the RSI event higher-ending percentage minus the All candles percentage. Direction, price change and movement charts: Actual measurements for each RSI group and All candles, shown separately. All candles has not been subtracted. Yearly chart: RSI event rows show differences from All candles (pp). All candles rows show actual higher-ending percentages (%). Actual higher, lower and unchanged percentages ten candles later. All candles is shown separately; its percentage has not been subtracted. “Ended higher” compares the tenth subsequent close with the starting close. For the shared-date, all-year cohort: 5m: first reach ≥70 ended higher in 44.53% of observations (N = 14,983); first reach ≤30 in 60.57% (N = 14,644). All candles: 50.94% (N = 1,433,175). 15m: first reach ≥70 ended higher in 47.09% (N = 5,802); first reach ≤30 in 59.25% (N = 5,347). All candles: 51.45% (N = 487,903). 1h: first reach ≥70 ended higher in 51.11% (N = 1,980); first reach ≤30 in 53.03% (N = 1,650). All candles: 52.36% (N = 124,109). 4h: first reach ≥70 ended higher in 58.59% (N = 640); first reach ≤30 in 58.11% (N = 444). All candles: 53.78% (N = 31,247). Different elapsed horizons prevent ranking the timeframes. █ What do the percentage-point differences mean? The main chart at the top of this Idea subtracts the matching All candles higher-ending percentage from each event group's percentage. For 5m, ten subsequent observed candles, shared dates and all years combined, first reach ≥70 ended higher in 44.53% (N = 14,983), versus 50.94% (N = 1,433,175) for All candles. The difference is −6.41 percentage points, calculated from unrounded counts. This compares ending frequencies, not return sizes. Lower and unchanged endings stay in both totals; All candles includes event closes. At higher overbought and lower oversold thresholds, lower endings after overbought events and higher endings after oversold events were not consistently more frequent. Ending direction alone does not establish a price reversal. █ How RSI events were counted An overbought episode begins when RSI moves from below 70 to at least 70 and resets below 70. An oversold episode begins when RSI moves from above 30 to at most 30 and resets above 30. Only the first reach of each threshold counts per episode: 70/75/80/85/90 and 30/25/20/15/10. Equality counts; a jump can reach several thresholds together. Staying beyond a threshold, or recrossing a deeper threshold before the episode resets, adds no new event. A new episode can count again. Each threshold uses its own qualifying close, so starting prices and dates can differ. These results do not describe every candle beyond a threshold. █ Can the average be positive when most endings are lower? Actual mean and median ending returns; All candles is shown separately, not subtracted. Open markers show means; filled markers show medians. Scales differ by timeframe. For 15m, ten subsequent candles, shared dates and all years combined: First reach ≥70, N = 5,802: mean ending return +0.0807%; median −0.0486%. The group ended lower in 52.72% of observations. All candles, N = 487,903: mean +0.0291%; median +0.0186%. It ended lower in 48.17%. First reach ≤30, N = 5,347: mean +0.1573%; median +0.1889%. It ended lower in 40.70%. What caught my attention in the 15m ≥70 group was that it ended higher less often than All candles, yet its mean return was higher. Most observations still ended lower, and the median was negative. Direction counts endings; the mean also reflects return sizes; the median is their middle value. On 4h, ten candles, shared dates and all years combined, ≤30 (N = 444) ended higher more often than All candles (N = 31,247): 58.11% versus 53.78%. Yet its mean return was smaller: +0.2910% versus +0.4407%. An ending return still leaves out the highs and lows reached along the way. █ How much did price move inside the window? Actual separate medians of ending return, maximum upside and maximum downside; All candles is shown separately, not subtracted. Downside is a positive magnitude. Ending return is the percentage change from the starting close to the tenth subsequent close. Maximum upside and downside use subsequent highs and lows relative to the starting close, excluding the trigger candle's high and low. Downside is a positive magnitude. For 15m, first reach ≤30, ten subsequent candles, shared dates and all years combined, N = 5,347, the median ending return was +0.1889%. The separate median maximum upside was 0.7482%, and median maximum downside was 0.6915%. For the matching 15m All candles comparison, N = 487,903, the respective values were +0.0186%, 0.4407% and 0.4764%. Both median maximum upside and median maximum downside were larger for the 15m ≤30 group than for All candles. These are separate medians, so they do not describe one typical price path or show which move happened first. All three medians use each group’s full sample, regardless of ending direction. Chart scales differ by timeframe and measurement. █ Sample size matters For 4h, first reach ≤10, ten candles, shared dates and all years combined, 88.89% ended higher, from just N = 9 observations. Matching All candles: 53.78%, N = 31,247. With N = 9, changing one outcome between higher and not higher shifts the percentage by about 11 percentage points. █ How did the results vary by year? RSI event rows: higher-ending percentage minus the matching All candles percentage, in percentage points (pp). All candles rows: actual higher-ending percentages (%). The pooled column includes partial years. RSI event cells show the event group's higher-ending percentage minus its matching annual All candles percentage. Positive cells do not mean positive returns. The right-hand column pools the shared interval, rather than averaging yearly differences. In the complete years shown, 5m ≥70 is below its annual comparison every year, while 5m and 15m ≤30 are above theirs. The 1h and 4h groups have mixed signs across years. Read the sample counts alongside the colors. In UTC trigger year 2025, under the same 5m, shared-date, ten-candle rules, first reach ≤30 ended higher in 50.85% (N = 1,349), versus 50.21% for that year's All candles (N = 105,084): +0.64 percentage points. Mean returns were −0.0246% and +0.0003%, respectively. Splitting the same data by year does not establish a trend or explain the differences. The full annual tables retain partial years, small samples and empty groups, whose percentages are undefined. █ Research notes and limits Interim research snapshot: October 11, 2026. The 5m, 15m, 1h and 4h datasets passed the coder’s automated data-validation checks. Shared eligible starting closes begin on June 20, 2012 at 20:00 UTC. Eligible starting candles across all four timeframes run through October 8, 2026 at 12:00 UTC. The charts measure the next ten candles, while eligibility requires twenty complete future candles. Confirmed data extends through October 11, 2026 at 21:00 UTC (20:00 UTC for 4h). Maximum-history results remain separate. Years use the UTC trigger close; windows can cross year boundaries. Eligibility excludes 1,000 initial loaded candles. Episodes already active before eligibility are suppressed until reset; fresh boundary crossings count. The main chart’s 2M+ sums the four shared-date, all-year, ten-candle All candles counts (2,076,434). TradingView’s Deep Backtesting made this historical coverage possible. These counts include overlapping windows and the same history viewed at different candle durations. Observations are not independent: windows overlap, thresholds can share an episode, and event closes also appear in All candles. Gaps and inconsistent open/high/low/close candles were preserved and flagged. No cleaned-data comparison was supplied. The refreshed summaries matched the original-period results within numerical tolerance. The coder reports that two historical 5m candles differed in open/low fields between the original and refreshed snapshots, without changes to measured outcomes. The cause remains unconfirmed. A check against TradingView’s built-in RSI and tests with a longer warm-up remain outstanding. Some original warm-up RSI and episode fields were unavailable for direct comparison. No uncertainty estimates are supplied. This post selects examples from the broader report. Qualifying RSI events were associated with different historical outcomes across the studied timeframes and horizons. Higher-ending frequency, average return and movement within the window described different aspects of those outcomes. These descriptive findings do not establish statistical significance, predictive performance, a trading edge or profitability. The observed frequencies should not be treated as probabilities for the next RSI event.