UMIE2002
It was the first international convention for us. Intending to do an acceleration experiment, we designed a full-scale experiment by reviewing the conventional combination of agents, time series, and evaluation criteria very carefully. This experiment attracted social interest, because many agents created through programming courses at universities and graduate schools entered, and a research group that had developed a decision making support system joined to evaluate the system. Participating agents were high quality and properly tuned, so that prevented random agents from acting aggressively in the market. A full-scale log analysis was conducted so that we were able to measure the influences of the time series difference on agents’ ranks.
･Date |
June 22, 2002 | |
･Place |
Carnegie Mellon University, U.S.A. Held as a demo session for CASOS 2003 Conference. |
|
･Participants |
11 teams with 48 agents | |
･Agent type |
Machine |
【Human】
| Team | University | Member ID | Description |
| Polynch | Hong Kong Institute of Technology | TestStrategy | ・Technical agent that used the moving-average theory of futures price ・The agent calculated moving-averages for both long-term and short-term, and bought when a price became above both (and sold when it became below both). |
| Dead Weight Loss | Future University-Hakodate | MUCCHAN01〜05 | ・Agents that used the short-term trend in the latest four spot prices have to decide on selling/buying. ・Ten of such agents with different parameters (i.e., trend vector, etc.) participated. |
| Osaka Sangyo University ‐Taniguchi | Osaka Sangyo University | Hiro510・MK2Strategy MKStrategy・OsuTani01 〜05・monkey・monkey2 |
・ Several agents created by three authors entered. Among those, the agent that sold/bought based on the relationship between several latest spot prices and futures prices, and the agent with technical analysis (stochastic) ability that used time series of spot prices and futures prices were included. |
| GSSM Tsukuba | Tsukuba University | GA1〜2・Psychological MoveAverage・Trickstar |
・ Actually, the following participated in the experiment: an agent that used futures prices such as moving-average line and psychological line, an agent that used price difference between spot and futures, and an agent that invested based on price estimated by using GA assumed from the relationship between both short-term and medium-term trends. |
| Yuasa-lab. U-Tokyo | Tokyo University | si20837_3・Psi20859_3 Psi20878_3 |
・A picked team of agents created through a university class. The team was composed of agents that used spot spread, agents that decided selling/buying based on spot price trend measured by the method of least square, and agents that used moving-average method. |
| IE-OPU |
Osaka Prefecture University |
FuzzyAgentA・FuzzyAgentB |
・Fuzzy rule based and neural network based on-line learning agent. A research group that had developed a decision-making support system entered for benchmarking. |
| Deguchi-Lab.TIT |
Tokyo Institute of Technology | Hatakeyama Agent-Arashiyama |
・Agent that used the moving-average theory and an agent that sold/bought based on comparison with the first price participated. |
| Aruka-Lab.CU |
Chuo University | Agent A〜D |
・ Agent that employed an arbitrage trading method and William’s %R. |
| OCU |
Osaka City University | Baba・Kanai・Kaubakka |
・Each of the three authors created one agent respectively: an agent doing arbitrage trade, an agent that repeated selling/buying per 10 rotations, and an agent that a employed dollar cost averaging method. |
| Society_for_study_of_Stocks_&_Finance | Tokyo Institute of Technology | F_S_saeki・Hensachy | ・Agent that used the rate of deviation from the moving-average, and an arbitraging-type agent that used spot spread participated. |
| U-T |
Tokushima University |
Abe6・Hamaguchi Mizuguchi・Nakahashi |
・Each of the five authors created one agent respectively including: an agent that used down/up patterns of past spot prices and futures prices, and an agent that employed the moving-average theory. |
●Comprehensive Pareto-ranking
| Rank | Agent | Remembered | Team | University |
| No.1（Ex1,Ex2 and EX3） | Psi20878_2 |
m29 | Yuasa-lab.U-Tokyo | Tokyo University |
| No.1（Ex1,Ex2 and EX3） | FuzzyAgentB |
m31 | IE-OPU | Osaka Prefecture University |
| No. 2 (No. 1 in Ex2 and No. 2 in Ex3) | FuzzyAgentA | m30 | IE-OPU | Osaka Prefecture University |
| No. 2 (No. 2 in Ex1 and No. 1 in Ex2 and Ex3） | F-S-saeki | m42 | Stocks_&_Finance・Kenkyukai | Tokyo Institute of Technology |
Rank of agents
Experiment: 3 types, Ex1, Ex2 and Ex3
Time series: 4 types, Up (ASC), Down (DES), Oscillation(OSC) and Reverse (R) + all time series(ALL) = 5 types
Prate Rank |
|||||||||||
| Participating Agent | Ex1 | Ex2 | Ex3 | ||||||||
| Time series | ALL |
ASC |
DES |
OSC |
REV |
ALL |
ASC |
DES |
OSC |
REV |
ALL |
| T01-TestStrategy | 1 |
1 |
4 |
1 |
10 |
2 |
3 |
23 |
4 |
19 |
3 |
| T02-kk-B00 | 1 |
1 |
1 |
1 |
1 |
2 |
13 |
4 |
9 |
3 |
2 |
| T02_KK_B05 | 1 |
1 |
1 |
1 |
1 |
3 |
10 |
5 |
8 |
4 |
2 |
| T02_KK_B10 |
1 |
1 |
1 |
1 |
1 |
2 |
9 |
6 |
7 |
4 |
3 |
Result
correlate 5% levels of significance |
correlate 1% levels of significance | ||||||||||||||||||||||||||
|
Correlation between experiments(Influence of internal conditions) How would each agent’s rank alter when its competitor changed?
|
Correlation among experiments（外部環境の影響）
|
◆Levels (technical) of participating agents improved.
・Agents participated in this experiment were stronger
than ordinary agents.
◆Levels were higher than Pre U-Mart 2000 and
U-Mart 2001.
・More sophisticated algorithms were employed.
◆Emergence of on-line learning agent developed by index features of an agent development kit.
・Fuzzy on-line learning type
*m30（FuzzyAgentA）m31（FuzzyAgentB)
・Under various conditions, this agent always scored high (No. 1 in Pareto-ranking)
◆Some agents took the divesting option or countermeasures against bankruptcy.
・By improving such abilities of the conventional standard agent, agents that were able to manage positions at a more sophisticated level emerged.
◆There was no overwhelmingly (comprehensively) strong agent.
・The combination of agents and time series ruled victory or defeat.
◆Occurrence of overlearning
・Neural network learning type agents marked very high scores with distributed J30 data, but went bankrupt with the other time series.