Aronsson Folkesson 2023
“Stock market analysis with a Markovian approach: Properties and prediction of OMXS30” is a first-cycle (bachelor-level) degree project, 15 credits, in Applied Mathematics and Industrial Economics at KTH Royal Institute of Technology, Stockholm, written by Max Aronsson and Anna Folkesson under supervisors Björn Wehlin and Hans Lööf. It is a student dissertation, not a peer-reviewed journal article — a status worth flagging when weighing its evidence. The study has two halves: a descriptive analysis of the OMXS30 Index as a Markov chain (transition matrix, steady-state distribution, hitting times) and a predictive evaluation of a Markov Chain Trading Model for next-day state forecasting. It appears in this vault as the clearest fully-worked example of how a discrete-state price-transition chain is constructed and how weakly it performs once held out of sample.
The methodology is careful for a BSc project. Daily OMXS30 closing prices from April 2020 to April 2023 are converted to percentage returns and discretised into six magnitude-bucket states D3-U3 (three “down” and three “up” buckets, split at ±0.5% and ±1%), plus an aggregated two-state up/down version. Transition probabilities are estimated by maximum-likelihood counting, P_ij = n_ij / n_i, and re-estimated over a sliding window so the matrix tracks recent dynamics. Window size is tuned by blocked time-series cross-validation — a genuine attempt to avoid data leakage and overfitting. Because no single window size dominated, the final predictor is a voting ensemble of ten chains with window sizes drawn randomly (with replacement) from 5-50 days. The model is treated as a classifier and scored on a held-out test set (January–April 2023) via confusion-matrix accuracy, precision, recall and F1.
Aronsson Folkesson 2023 [proposes_model] Markov Chain Trading Model Aronsson Folkesson 2023 [tests-on] OMXS30 Index Aronsson Folkesson 2023 [defines] State-Count Selection
The headline result is deflationary. A single Markov chain predicted next-day movement no better than random chance. The ten-chain ensemble with the six-state configuration reached an out-of-sample accuracy of 17.1% versus the 16.7% random benchmark — a gap of roughly 0.4 percentage points. The second-order six-state ensemble reached 18.7%, again only marginally above benchmark. The aggregated up/down model is worse: the first-order version scored exactly 50.0% (the random benchmark) and the second-order up/down version scored 42.7%, below random chance. Crucially, second-order chains did not beat first-order despite using more information — the extra states simply spread thin data thinner, exactly the State-Count Selection / data-sparsity trade-off the authors anticipate. The authors are candid: “it is not reasonable to expect that the simple prediction approach with Markov chains would considerably outperform random chance,” and they warn the up/down model carries a positive bias (it predicts “up” too readily) that is “not suitable” for a risk-averse investor.
Aronsson Folkesson 2023 [compares_benchmark] Random Walk Benchmark Aronsson Folkesson 2023 [supports] State-Count Selection Aronsson Folkesson 2023 [contradicts] Markov Chain Trading Model
For grading purposes the gaps are decisive. There is no trading backtest at all — the model is scored on classification accuracy, never converted to a P&L. There are therefore no transaction costs, no slippage, no Sharpe, no drawdown, and no buy-and-hold comparison. The benchmark cleared is an error/accuracy metric (random chance), not a costed financial baseline. The test sample is a single four-month window of 2023 data, far too small to support a profitability claim, and no code or replication package is provided. A 0.4-percentage-point accuracy edge over random chance would be comfortably erased by realistic Transaction Costs and Slippage. The descriptive half of the paper is more durable — the steady-state distribution and hitting times genuinely characterise OMXS30’s structure (volatility clustering, mean reversion within ~5 days, mild long-run upward drift) — but that is Regime Classification-style description, not a tradeable edge. Net grade: weak — out-of-sample but no costs, no benchmark beyond random chance, tiny sample, no backtest, no replication.
Transaction Costs and Slippage [contradicts] Aronsson Folkesson 2023 Aronsson Folkesson 2023 [relates] Regime Classification
Connections
- Markov Chain Trading Model — proposes_model, 2023, source: https://kth.diva-portal.org/smash/get/diva2:1823899/FULLTEXT01.pdf
- OMXS30 Index — trades_market, 2020-2023, source: https://kth.diva-portal.org/smash/get/diva2:1823899/FULLTEXT01.pdf
- Random Walk Benchmark — compares_benchmark, 2023, source: https://kth.diva-portal.org/smash/get/diva2:1823899/FULLTEXT01.pdf
- State-Count Selection — suffers_overfitting_risk, 2023, source: https://kth.diva-portal.org/smash/get/diva2:1823899/FULLTEXT01.pdf
- Non-Stationary Transition Matrix — suffers_overfitting_risk, 2023, source: https://kth.diva-portal.org/smash/get/diva2:1823899/FULLTEXT01.pdf
- Transaction Costs and Slippage — excludes_costs, 2023, source: https://kth.diva-portal.org/smash/get/diva2:1823899/FULLTEXT01.pdf
- Out-of-Sample Backtesting — lacks_live_evidence, 2023, source: https://kth.diva-portal.org/smash/get/diva2:1823899/FULLTEXT01.pdf
- Regime Classification — detects_regime, 2023, source: https://kth.diva-portal.org/smash/get/diva2:1823899/FULLTEXT01.pdf