AlphaFlow Quant Research: Low-Volatility Upward Drift Breakdown Study (1H)

Objective

Design and execute a rigorous quantitative study to determine how often a small, low-volatility, upward-trending consolidation breaks down rather than breaks upward, particularly when it occurs after a sharp bearish impulse.

The pattern of interest is a short, quiet, gently rising price structure. Price drifts upward in a narrow range with relatively low volatility, but the recovery may be weak and vulnerable to another downward move.

This study is about measuring the conditional probability of breakdown, not assuming the pattern is bearish.

The primary research question is:

Following a bearish impulse, does a low-volatility upward drift have a statistically elevated probability of breaking down before breaking upward, compared with appropriate baseline conditions?

Do not assume the hypothesis is true. The experiment must be capable of falsifying it.

1. Inspect the existing infrastructure first

Before implementing anything:

1. Inspect the existing AlphaFlow research framework, historical data loaders, feature calculations, experiment runners, event-study utilities, backtesting infrastructure, and reporting conventions.
2. Identify reusable components and extend existing abstractions rather than creating duplicate infrastructure.
3. Inspect the existing 1H candle dataset and verify timestamp alignment, missing candles, duplicate records, symbol coverage, and data quality.
4. Confirm the actual available history and whether sufficient 1-minute OHLCV data exists for resolving ambiguous intrabar boundary touches.
5. Document the implementation plan and data limitations before running the experiment.

Do not invent APIs, datasets, existing functions, or results. If something is unavailable, explicitly report it.

2. Dataset and universe

Use the established AlphaFlow research universe:

• BTCUSDT
• ETHUSDT
• SOLUSDT
• ZECUSDT
• NEARUSDT
• XRPUSDT
• OGNUSDT
• HYPEUSDT
• BNBUSDT
• DOGEUSDT

Venue: Binance USD-M perpetual futures.

Timeframe: 1-hour candles, preferably resampled from 1-minute data using the existing pipeline.

Research period: January 2025 through September 2026, subject to actual data availability.

Use the same universe, data conventions, and infrastructure as previous AlphaFlow experiments so results remain comparable.

3. Define the pattern mathematically

Represent the pattern using causal features computed at the close of each 1H candle.

Do not use chart-image recognition. The goal is a reproducible numerical definition.

A. Prior bearish impulse

Create separate cohorts:

• Cohort A: upward-drift patterns preceded by a bearish impulse.
• Cohort B: upward-drift patterns without a qualifying prior bearish impulse.

Test alternative impulse definitions using:

• Returns over the preceding 3, 6, and 12 hours.
• Return normalized by rolling realized volatility.
• Decline from the preceding rolling high.
• ATR-normalized price displacement.

Test multiple impulse thresholds rather than selecting one threshold based on final performance.

Ensure the impulse occurs before the detected consolidation. Do not use future price data to define the impulse.

B. Upward drift

For rolling windows of 4, 6, 8, 10, and 12 hourly candles, calculate:

• Linear-regression slope of log-price.
• Slope normalized by rolling volatility or ATR.
• Cumulative close-to-close return.
• Percentage of positive candles.
• Number of higher closes.
• Number of higher swing lows, using a causal swing definition.
• Total high-low range normalized by ATR.
• Maximum drawdown within the window.

Require positive drift, but test alternative minimum and maximum slope thresholds.

The purpose is to identify a gradual ascent, not a strong directional breakout.

C. Low volatility and range compression

Calculate:

• Realized volatility over the pattern window.
• ATR over the pattern window.
• Ratio of short-term ATR to a longer-term ATR baseline.
• Mean and median true range relative to the historical baseline.
• High-low range divided by ATR.
• Candle body size relative to total candle range.
• Volume relative to its trailing average.

Test several compression definitions independently.

Do not equate low volume with selling exhaustion. Volume is a contextual feature, not a mandatory assumption about market intent.

D. Weak recovery characteristics

Calculate:

• Distance from the fast and slow EMAs.
• Normalized EMA spread.
• Fast EMA ROC.
• Change in EMA ROC (acceleration).
• EMA spread velocity.
• Distance from the preceding impulse low.
• Distance from the consolidation high and low.

These features should initially be descriptive or used for ablation tests. Do not make every feature a mandatory condition for pattern detection.

E. Consolidation boundaries

For each detected event, freeze the pattern high and low at the end of the qualifying window:

• pattern_high
• pattern_low
• pattern_midpoint
• pattern_range
• pattern_duration

The boundaries must not move after the event has been detected.

Avoid look-ahead bias when determining when a pattern starts and ends.

4. Primary outcome: which boundary breaks first?

At the close of the detection candle, define the next candle as the beginning of the outcome window.

Evaluate forward horizons of:

• 3 hours
• 6 hours
• 12 hours
• 24 hours
• 48 hours

The primary horizon is 12 hours.

For each event, determine whether price reaches either frozen boundary first.

Outcome classes

1. BREAKDOWN_FIRST: price trades below pattern_low before reaching pattern_high.
2. BREAKOUT_FIRST: price trades above pattern_high before reaching pattern_low.
3. NEITHER: neither boundary is touched within the specified horizon.
4. AMBIGUOUS_SAME_BAR: both boundaries are touched in the same 1H candle and the order cannot be established from hourly OHLC data.
5. DATA_INVALID: insufficient or invalid future data.

Use 1-minute data to resolve same-hour boundary ordering when available. Otherwise retain an ambiguous classification.

Do not arbitrarily classify ambiguous events as breakdowns or breakouts.

Report the following separately:

• Breakdown-first probability among resolved events.
• Breakout-first probability among resolved events.
• Neither-boundary probability.
• Ambiguous-event frequency.
• Number of total and resolved events.

Calculate:

breakdown_first_rate = breakdown_first_count / resolved_event_count

where resolved events include breakdown-first and breakout-first outcomes. Also report the rate using all valid events as the denominator so censoring and unresolved outcomes remain visible.

5. Critical distinction: distinguish pattern breakdown from bearish profitability

Breaking the pattern low does not necessarily mean the trade is profitable.

For each event, also calculate:

• Forward returns over 3, 6, 12, 24, and 48 hours.
• Maximum favorable excursion (MFE).
• Maximum adverse excursion (MAE).
• Time to first breakdown.
• Time to first breakout.
• Post-breakdown return over subsequent horizons.
• Post-breakout return over subsequent horizons.
• Whether the initial break reverses quickly.

Use consistent definitions and normalize excursion measures by ATR where useful.

This separates three questions:

1. Does the pattern break down more frequently?
2. When it breaks down, does price continue falling?
3. Can that continuation be captured by a realistic trading strategy?

The first is the primary question. The others are secondary.

6. Baseline comparisons

A breakdown rate alone is not evidence of predictive value.

Compare the detected pattern against:

Baseline A — Random comparable windows

• Windows matched by symbol, volatility regime, duration, and market period.
• Match prior bearish impulse status when applicable.

Baseline B — Low-volatility sideways consolidation

• Similar volatility compression and duration.
• No positive-drift requirement.

Baseline C — Upward drift without low volatility

• Similar positive slope and duration.
• No volatility-compression requirement.

Baseline D — Upward drift after a bearish impulse

• Same prior impulse cohort but without the low-volatility requirement.

Baseline E — All eligible post-impulse consolidations

• General post-selloff consolidation behavior.

Ensure matching does not use future outcomes.

For each comparison, report the absolute difference in breakdown probability, relative difference where meaningful, confidence intervals, and event counts.

The central question is whether low-volatility upward drift adds information beyond a generic consolidation or a post-selloff environment.

7. Feature ablation

Run controlled ablations to isolate the contribution of each condition:

1. Positive drift only.
2. Low volatility only.
3. Positive drift + low volatility.
4. Positive drift + low volatility + prior bearish impulse.
5. Add weak-recovery characteristics.
6. Add EMA ROC and acceleration.
7. Add EMA spread velocity.
8. Add volume features.

Measure incremental changes in:

• Breakdown-first probability.
• Breakout-first probability.
• Event frequency.
• Confidence intervals.
• Forward returns.
• MFE and MAE.

Do not assume that more conditions produce a better result. Record whether a feature improves conditional prediction, merely reduces the sample size, or adds no measurable information.

8. Parameter sensitivity

Systematically test:

• Pattern duration: 4, 6, 8, 10, 12 bars.
• Outcome horizons: 3, 6, 12, 24, 48 hours.
• Multiple positive-slope thresholds.
• Multiple volatility-compression thresholds.
• Multiple impulse thresholds.
• Alternative ATR and realized-volatility lookbacks.

Use a predefined parameter grid.

Report the complete experiment grid or a machine-readable equivalent, not only the best-performing configuration.

Account for multiple testing. Do not select and report the most favorable combination as though it were the original hypothesis.

9. Statistical validation

Use:

• Confidence intervals for breakdown and breakout probabilities.
• Bootstrap methods appropriate for serially dependent financial observations.
• Block bootstrap or equivalent dependence-aware methods where appropriate.
• Chronological train/validation/test splits.
• Walk-forward validation.
• Separate results by symbol.
• Separate results by volatility and broader market regimes.

Overlapping pattern windows can create highly dependent events. Deduplicate or cluster overlapping events for appropriate analyses, and report both raw event counts and independent-event estimates where feasible.

Freeze the final pattern definition before evaluating the untouched out-of-sample period.

10. Optional trading-strategy evaluation

Only after the event study is complete, evaluate a short strategy triggered by a breakdown below the pattern low.

For this secondary test:

• Signal becomes available only after the relevant candle closes.
• Execute at the next available candle open or a more realistic simulated fill.
• Include transaction costs, slippage, funding, and position reversals.
• Use 6 bps per side as the baseline cost assumption.
• Run cost sensitivity at 0, 3, 6, 12, and 20 bps per side.
• Prevent future data from influencing the entry, stop, or exit.
• Specify stop-loss, take-profit, and time-based exits before comparing results.
• Compare against a simple breakdown strategy without the pattern filter.

Do not optimize strategy parameters on the final test set.

11. Required deliverables

Produce:

A. Research report

Include:

1. Executive summary.
2. Exact pattern definition.
3. Dataset and data-quality checks.
4. Total event counts.
5. Breakdown-first, breakout-first, neither, and ambiguous rates.
6. Results for each forward horizon.
7. Results with and without the preceding bearish impulse.
8. Baseline comparisons.
9. Feature-ablation results.
10. Parameter-sensitivity analysis.
11. Per-symbol and per-regime results.
12. Confidence intervals and statistical limitations.
13. Out-of-sample results.
14. Optional strategy results after realistic costs.
15. A clear conclusion: supported, unsupported, inconclusive, or requiring further validation.

B. Reproducible artifacts

Save versioned outputs in the existing AlphaFlow experiment structure:

• Experiment configuration.
• Pattern-detection implementation.
• Feature definitions.
• Event-level dataset.
• Aggregate results.
• Parameter-sensitivity results.
• Validation results.
• Report and visualizations.

Use existing repository conventions and avoid introducing unnecessary dependencies.

C. Useful visualizations

Generate:

• Example detected patterns overlaid on 1H price charts.
• Distribution of pattern durations and slopes.
• Breakdown-first probability by duration and horizon.
• Pattern versus baseline breakdown rates with confidence intervals.
• Forward-return distributions.
• Results split by symbol and market regime.
• Parameter-sensitivity heatmaps.

Include examples of false positives and patterns that break upward.

12. Research discipline

• Do not fabricate results.
• Do not change the hypothesis after inspecting the test results without explicitly versioning a new experiment.
• Do not use future information to identify or label the pattern.
• Do not interpret correlation as causation.
• Do not call a breakdown probability predictive unless it outperforms an appropriate baseline with adequate uncertainty analysis.
• Do not declare profitability based on win rate alone.
• Report sample sizes prominently.
• If a feature or data source is unavailable, report that limitation instead of silently substituting another.

Final question to answer

On the 1H timeframe, after a bearish impulse, how often does a low-volatility, gently upward-trending consolidation break below its low before breaking above its high within the next 12 hours, and is this probability materially higher than for comparable consolidations?

Answer this question first. Everything else should help explain whether the result is real, robust, and potentially tradable.

← All reports · source (.md)