Free lesson · Charts & patterns

Research review: what the evidence does and does not establish

Open interactive lesson Practice calculations Explore labs

Start with the idea

Evidence becomes useful when its test conditions are visible. A new publication can suggest a better experiment without establishing that a model will earn money in your market.

Symbols, units & horizon

Gross P&L: currency result before the explicitly listed expenses
Execution: fees plus spread/slippage, with no double counting
Borrow: securities-loan expense
Funding: financing expense
Net P&L: currency result after these deductions
Review date: date sources were checked, not a claim of automated monitoring

When and why to use this

Use this reading list when designing a pattern study and reviewing whether a proposed bot has learned incremental information rather than a data or evaluation artefact.

Literature checked 11 September 2026. This review searched recent primary academic sources alongside foundational work. It is a dated reading list, not a claim of exhaustive coverage or a live performance ranking. Publication status, sample choice and access limits matter when translating a result into a trading hypothesis.

Primary source & status	Useful finding	Limit on interpretation
Lo, Mamaysky & Wang (2000), Journal of Finance ↗	Computational pattern definitions and statistical comparisons show how to replace subjective labels with reproducible tests. Some patterns contained conditional return information in the studied historical US stock sample.	Historical evidence from 1962–1996 is not a current after-cost trading result. Full author manuscript reviewed.
Radfar (2025), peer-reviewed ↗	A study of 12 Tehran stocks shows how level-prediction accuracy can look strong relative to an inappropriate baseline. Compare with a constant-price prediction before interpreting a complex model.	Small, market-specific sample and model-selection choices limit generalisation. Full text reviewed; this does not prove all chart models fail.
Kim et al. (2025), CIKM / author preprint ↗	Shapelet-based directional forecasting uses smoothing, time alignment and learned subsequences to make chart-pattern research explicit.	The reported universe includes Bitcoin and three US stocks. Replicate causal preprocessing, costs and validation before treating it as an investment result. Full preprint reviewed.
Azevedo, Hoegner & Velikov, AEA 2025 conference paper ↗	Machine-learning strategy results depend materially on publication timing, transaction costs and the evaluation period. Use this as a reason to test implementation realism.	A working-paper analysis of particular anomaly strategies does not rank every chart or bot method. Conference manuscript reviewed.
Bøjstrup, Veliyev & Wulff (July 2026), working paper ↗	The abstract decomposes chart-image prediction advantages and reports that explicit features can recover much of the remaining chart-specific signal.	Under review according to the author’s page, not an accepted journal result. Only abstract and metadata reviewed; full PDF access unavailable. Do not treat its reported results as independently verified.

The practical teaching choice is to begin with interpretable OHLCV variables, simple causal baselines and realistic costs, then require more complex models to show incremental value. This is our synthesis of the sources, not an academic guarantee of the best-performing architecture. Older mechanics are included for their explanatory value; new research informs how to evaluate them.

Research update protocol: search recent papers by concept and publication date, then read primary methods and limitations.
Record data dates, markets, sample size, publication status, access level, execution assumptions and whether the result concerns prediction or traded profit.
Reproduce the baseline, information timing and net-cost evaluation before importing a performance claim.
Keep the rejected and unsuccessful trials so the final study reflects the full search, not only the best chart.

Advanced recognition update · 12 September 2026

Kim et al. · CIKM 2025 · shapelet forecasting ↗

Full methods, data split and trading protocol reviewed 12 September 2026. BTC/USD history 2014–May 2025; AAPL, BRK.B and XOM history 2008–May 2025, with distinct chronological splits. The paper uses roughly 20-day motifs and four-day trades with .1% fees on each side. Small universe, alignment choices and retrospective confidence ranking limit direct deployment. Our banded-DTW example is not a replication.

Hu et al. · April 2026 preprint · multi-scale VLM benchmark ↗

Full data-construction and evaluation sections reviewed 12 September 2026. The reported US/China OHLCV dataset spans 2015–2025, with daily/weekly charts and a main 30-day target. Regime and horizon diagnostics motivate matched numeric baselines. Historical pretraining exposure, selected datasets and prediction-focused evaluation limit claims about executable profit.

Wang · June 2026 preprint · controlled candlestick audits ↗

Benchmark construction and limitations reviewed 12 September 2026. Controlled shadow markets use null outcomes, altered local evidence and trend/label swaps, rather than a real-market trading sample. The study motivates checking trend and rendering shortcuts. Its synthetic mechanism and evaluated frozen models do not establish performance for another market or model.

Nguyen et al. · Information Sciences 2026 · VARDiff author repository ↗

Author README and publication citation reviewed 12 September 2026; publisher full text unavailable. The repository describes visual retrieval guiding diffusion forecasts on nine stock datasets. Exact sample dates and execution assumptions were not independently verified. Retrieve only historical neighbors whose required continuation is already known; a forecast-error claim is not a net-profit claim.

Research sources, review dates and limitations

Extend the research question

Record how bars are constructed and when each pattern becomes observable. Compare a visual marker’s plotted time with its earliest legal decision time.

Continue with the connected research module →

Accounting identity

Write a reproducible net-result bridge

Choose one untouched evaluation period and a fixed position rule. Sum its recorded gross currency trade P&Ls.
Sum explicit fees, spread/slippage, borrow and financing over that same ledger. Subtract each exactly once; divide by a documented capital base only when that return convention is appropriate.

Work it by hand

If gross P&L=$1,000, execution=$400, borrow=$100 and funding=$150, net P&L=$350. Prediction accuracy alone cannot produce that ledger.

Python implementation

Self-contained teaching example. Python 3.10+; dependencies and input conventions are shown in the code and notation. Run in your own Python environment.

def research_pnl_bridge(gross_pnl, execution_cost, borrow_cost, funding_cost):
    """Use one consistent ledger and avoid counting expenses twice."""
    return gross_pnl-execution_cost-borrow_cost-funding_cost

print(research_pnl_bridge(1000,400,100,150))

Continue learning

Candles, Structures & Pattern Research — all lessons

Quantitative finance and development glossary · Python resources and libraries · Research sources and limitations