We built 19 Borsa Istanbul research systems but still cannot prove a real trading edge — what are we missing?

[sedatguner2000-coder]

We have been developing an independent decision-support system for Borsa Istanbul for approximately 22 months.

The system uses only delayed market information that was available at the exact decision time. It does not use future data, real-time private feeds, other stock markets, or automated order execution. It does not manage client money. Its purpose is to examine the market, identify candidates, compare their relative strength and risk, and produce a manual daily decision.

How the system works

Over time, we built and tested 19 different research architectures. These were not simple parameter changes; they examined different combinations of price behaviour, intraday movement, market and sector conditions, historical similarities, candidate ranking, downside protection, capital allocation, holding periods, and exit decisions.

The current AVCI architecture contains:

• 20 main modules
• 23 supporting files
• Data collection and validation
• Decision-time locking
• Market and sector analysis
• Proprietary pattern and similarity layers
• Candidate-pool creation
• Candidate filtering and ranking
• Risk and uncertainty checks
• Decision recording
• Next-session result and cost measurement
• Final consistency and evidence checks

Historical daily and real one-minute BIST price and trading data have also been examined. Information created after the decision time is not supposed to enter the decision process.

The unresolved problem

Despite thousands of hypotheses, simulations, tests, and several complete architecture changes, we have not been able to prove a repeatable and executable after-cost edge.

Promising historical results often weaken or disappear when:

• The market period changes
• Transaction costs are included
• Capital is distributed among candidates
• Holding and replacement rules change
• Large winning days or stocks are removed
• Previously unused periods are tested
• Realistic execution assumptions are applied

A further problem is that much of the historical period has already been examined during research. After thousands of experiments, even an apparently excellent historical result may simply be a false discovery caused by overfitting and repeated testing.

There are also unresolved differences between a paper result and actual capital growth. A correct candidate does not automatically mean a profitable trade. Entry price, liquidity, slippage, tradable quantity, transaction costs, corporate actions, holding time, and exit timing can all change the result.

Our historical records also do not provide a complete real-money ledger containing every decision, executed quantity, entry, exit, cost, and daily capital change. For this reason, some old results cannot be reconstructed as genuine executable performance.

At present, we cannot confidently distinguish between three possibilities:

1. Delayed BIST information may contain no persistent edge strong enough to survive all costs.
2. A real edge may exist, but it may be lost during ranking, allocation, holding, replacement, or risk control.
3. The apparent edge may exist only on paper because execution and historical validation assumptions are unrealistic.

What we are looking for

We are not looking for:

• Stock tips
• A ready-made strategy
• Another generic machine-learning model
• Hundreds of new indicators
• A twentieth architecture built without first identifying the real problem

We are looking for experienced researchers, graduate students, quantitative developers, market-microstructure specialists, or independent practitioners who are willing to examine this problem carefully and patiently.

The central question is:

What immutable and pre-registered test could determine whether AVCI contains a real, executable and persistent after-cost edge—or whether the apparent historical advantage is only the result of overfitting, repeated testing, or unrealistic execution assumptions?

We are especially interested in people with experience in:

• Backtest overfitting and false discovery
• Delayed financial data
• Emerging or relatively illiquid equity markets
• Borsa Istanbul
• Transaction costs, liquidity and slippage
• Portfolio allocation and risk
• Holding and replacement decisions
• Walk-forward or prospective testing
• Pre-registration and independent validation

A negative conclusion is acceptable. The objective is not to make an unsuccessful system appear successful. The objective is to determine, with defensible evidence, whether a real edge exists, where it disappears, or why it cannot be extracted under the current constraints.

This is not a quick question that can be solved with one indicator or a few comments. We are looking for serious contributors who are willing to understand the architecture and help define a small number of decisive experiments.

A concise anonymized technical summary can first be shared with serious contributors. Proprietary selection rules, the complete source code, and raw data that we do not have the right to redistribute will not be posted publicly.

Our core question is:

Why, although this large research infrastructure appears able to identify strong candidates, can we not convert that ability into repeatable, executable, after-cost capital growth across different market periods?

[yaroslavmak1995-prog]

Really appreciate the honesty in this post. The "promising results disappear when you add real costs" problem is something we encountered directly.

We built a 9-factor trend visualization indicator (Ocean Wave) and ran a 5-year backtest across 5 tickers (AAPL, TSLA, SPY, GOOGL, NVDA — 5,925 trading days). The results were sobering:

• Average win rate: 45.32% (below random 50%)
• High confidence did NOT improve accuracy (47.48% vs 42.64% — only 5pp difference)
• Uptrend predictions slightly better at 52%, but still not reliable enough to trade on
• GOOGL confidence swung 0% to 87% in a single day during volatile periods

The key insight from your post that resonates: "promising historical results often weaken or disappear when transaction costs are included." We saw the same — what looked like "edge" in raw win rate disappeared when you factored in the cost of acting on every signal.

What we learned: The value isn't in prediction (we can't predict markets). The value is in visualization — helping traders see 9 indicators at a glance instead of 9 separate charts. It's a decision-support tool, not a signal service.

I think the trading software industry has too much "97% win rate" marketing and not enough "we tried, here's what we found, here's what it's actually good for." We were honest about 45.32% win rate in our README with a badge, rather than hiding it.

Would be curious if you've considered pivoting from "find the edge" to "build the best decision-support tool" — accepting that edge is hard, but clarity of information has value too.

Repo: https://github.com/yaroslavmak1995-prog/ocean-wave-indicator

[sedatguner2000-coder]

Evet sanırım artık ben de vazgeçtim gerçekten çok yoruldum çok araştırdım binlerce hipotez simülasyon onlarca defa sistemin baştan kurmak aylarca inanılmaz yoğun çalışma gerçekten beni çok yordu artık ben de bunun tamamlanabileceğine olan inancimi kaybettim Ama bizim sorunumuz biz gecikmeli veri ile bir şeyler yapmaya çalıştık ihtiyacımız olan veri seti de çok yüksek fiyat olduğu için canlı veri ile gerçek veri ile çalışmayı düşünmedik belki gerçek canlı veri sistemi değiştirebilirdi ? veri kalitesini artırıp aradığımız EDGE’yi bulmamıza yardımcı olabilirdi belki Sizin söylemiş olduğunuz öneri ise bizim hayal ettiğimiz istediğimiz seviyede bir çalışma değil o sebeple çok üzgün olarak bu kadar emekler boşa gittiğini açıkça söyleyebilirim 😔😔 Nazik dönüşünüz geri bildiriniz ve tavsiyeniz için samimiyetle kalben en güzel duygularla 🇹🇷🇹🇷Türkiye’den teşekkür ediyorum keşke bir yol olsaydı ama olmadı iOS için Outlook<https://aka.ms/o0ukef> uygulamasını edinin

…

________________________________ Gönderen: yaroslavmak1995-prog ***@***.***> Gönderildi: Tuesday, 11 August 2026 19:03:18 Kime: PyneSys/pynecore ***@***.***> Bilgi: sedatguner2000-coder ***@***.***>; Author ***@***.***> Konu: Re: [PyneSys/pynecore] We built 19 Borsa Istanbul research systems but still cannot prove a real trading edge — what are we missing? (Discussion #73) Really appreciate the honesty in this post. The "promising results disappear when you add real costs" problem is something we encountered directly. We built a 9-factor trend visualization indicator (Ocean Wave) and ran a 5-year backtest across 5 tickers (AAPL, TSLA, SPY, GOOGL, NVDA — 5,925 trading days). The results were sobering: * Average win rate: 45.32% (below random 50%) * High confidence did NOT improve accuracy (47.48% vs 42.64% — only 5pp difference) * Uptrend predictions slightly better at 52%, but still not reliable enough to trade on * GOOGL confidence swung 0% to 87% in a single day during volatile periods The key insight from your post that resonates: "promising historical results often weaken or disappear when transaction costs are included." We saw the same — what looked like "edge" in raw win rate disappeared when you factored in the cost of acting on every signal. What we learned: The value isn't in prediction (we can't predict markets). The value is in visualization — helping traders see 9 indicators at a glance instead of 9 separate charts. It's a decision-support tool, not a signal service. I think the trading software industry has too much "97% win rate" marketing and not enough "we tried, here's what we found, here's what it's actually good for." We were honest about 45.32% win rate in our README with a badge, rather than hiding it. Would be curious if you've considered pivoting from "find the edge" to "build the best decision-support tool" — accepting that edge is hard, but clarity of information has value too. Repo: https://github.com/yaroslavmak1995-prog/ocean-wave-indicator — Reply to this email directly, view it on GitHub<#73?email_source=notifications&email_token=B3NTTT67ULYNIXJRDQWEPR35JM7UNA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZZG44DGMBWUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVRTG633UMVZF6Y3MNFRWW#discussioncomment-17978306>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/B3NTTT6PAO4U3OWHPL2SSC35JM7UNAVCNFSNUABIKJSXA33TNF2G64TZHM4TCMJUGAYDOMBYHNCGS43DOVZXG2LPNY5TCMBVGIYDIOBQUF3AE>. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS<https://github.com/notifications/mobile/ios/B3NTTTZYQAEBQ7XYA4WVGID5JM7UNA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZZG44DGMBWUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVJTG633UMVZF62LPOM> and Android<https://github.com/notifications/mobile/android/B3NTTT23L6ZRYCAH4YZGHF35JM7UNA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZZG44DGMBWUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVZTG633UMVZF6YLOMRZG62LE>. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***>

[wallneradam]

I am the maintainer here. This is not really a PyneCore question, but I do not want to leave it without an answer.

I spent about five years on this. I never found a strategy I would put my own money into blindly. And I do not regret a single month of it, because I learned an enormous amount along the way.

That is the part I would offer you. I think this only works if the goal is not the money. If the money is the goal, then every negative result is a loss, and after 22 months you feel exactly what you are feeling now. If the goal is understanding, then a negative result is still a result, and the work was not wasted.

You did not waste it. Decision-time locking, cost measurement, data validation: most people never build any of that. They skip straight to the backtest that looks good. You have the part that is actually hard to build.

One practical thing I can offer, from my own experience rather than from theory. When I was training RL agents, a single split misled me badly. The training outcome depended heavily on where the cut landed, because different periods were different market regimes. What I ended up doing was splitting the data into many blocks and letting the agents draw from them randomly. If I were validating a strategy today, I would do the same: validate across several separate periods instead of one long block at the end, and look at how much the result varies between them, not just the average. Consistency across regimes tells you more than one good number.

I do not think better or more expensive data would have changed your outcome. That is not usually where the problem is.

If you come back to this someday, come back to it as something you enjoy doing, not as something that owes you a return. That is the only way I have found to keep going for years without it wearing you down.

[yaroslavmak1995-prog]

Thank you for this response. The five-year perspective is valuable. You are right that if the goal is money, every negative result is a loss. If the goal is understanding, every result is data.

Your point about regime-dependent validation is spot on. We ran one continuous 5-year block per ticker. We did not test how the algorithm performs across different market regimes (2020 crash, 2021 bull run, 2022 bear market, 2023 recovery). That is a blind spot in our backtest — the average win rate of 45.32% might mask regime-specific performance that is actually useful (or even worse than average).

We will add regime-segmented backtesting as a next step. If the algorithm is consistently 45% across all regimes, that is one story. If it is 55% in strong trends and 35% in choppy markets, that is a different (and more useful) story.

Appreciate you leaving a thoughtful answer on a discussion that is not even about PyneCore. That says a lot about the community you are building here.

[yaroslavmak1995-prog]

@sedatguner2000-coder Your message resonates deeply. 22 months of intense work — decision-time locking, cost measurement, data validation, rebuilding systems dozens of times — that is not wasted. Most people never build any of that. They skip to the backtest that looks good and never question it.

You built the hard part: the discipline to ask "does this actually work?" and accept the answer. That is rarer than you think.

The expensive data question is worth addressing. In our experience, better data did not change the outcome — the edge was not hiding in data quality. But you are right that delayed data limits certain strategies (especially intraday). The question is: what strategy type matches the data you can afford?

If you ever come back to this — not to find the edge, but to build tools that help you see context faster — that is a different kind of work. Less pressure, less disappointment, still valuable. We pivoted from "find the edge" to "build the best visualization" and it changed everything. Not because we found the edge, but because the goal became achievable.

From Türkiye to Ukraine, there are builders trying to solve hard problems with limited resources. That is worth something regardless of the outcome.